Dynamic intelligent evaluation system for learning behavior of students in medical colleges and universities integrating AI method

Abstract

This invention belongs to the field of medical education, specifically relating to a dynamic intelligent evaluation system for the learning behavior of medical students that integrates AI methods. The system includes: a multi-source data acquisition module for collecting learning behavior data; a data preprocessing and fusion module for data cleaning, standardization, and correlation; a multimodal fusion analysis module that uses a multimodal fusion model to map the learning behavior data into dynamic quantitative scores for the student's ability dimensions; and an evaluation and feedback module for generating dynamic profiles of individual and group abilities, knowledge reinforcement suggestions, and learning path suggestions based on the dynamic quantitative scores. This addresses the problems of evaluation lag and strong subjectivity inherent in traditional medical education evaluation methods.

Description

Technical Field

[0001] This invention belongs to the field of medical education, and in particular relates to a dynamic intelligent evaluation system for the learning behavior of medical college students that integrates AI methods. Background Technology

[0002] Student evaluation is a core component of medical education. Its fundamental purpose is to scientifically and accurately measure and guide medical students' knowledge acquisition, skill development, and professional ethics, providing a basis for decision-making regarding teaching quality improvement and personalized learning support. For medical and pharmaceutical colleges, students' learning behaviors are highly complex, spanning various scenarios such as theoretical teaching, laboratory operations, clinical clerkships, and simulation training. Traditional evaluation systems are insufficient to fully capture their multidimensional and dynamic ability development process. Therefore, constructing an evaluation system capable of deeply integrating multi-source data and achieving intelligent analysis and dynamic feedback is of great significance for improving the quality of medical talent training and achieving precision and modernization in education.

[0003] Currently, the evaluation system for medical students' learning behavior mainly relies on summative assessments, such as theoretical exams, lab report grading, and objective structured clinical skills exams. These methods have significant limitations: First, the evaluation is highly delayed, failing to provide immediate feedback during student learning and training, thus missing the optimal opportunity for teaching intervention. Second, the evaluation dimensions are singular, overemphasizing knowledge recall and isolated skill operations, making it difficult to comprehensively and objectively assess students' decision-making ability, operational fluency, and professional competence in complex clinical situations. Finally, the evaluation process is subjective, especially for practical operations, which heavily relies on the examiner's personal experience, making it difficult to guarantee the consistency and fairness of evaluation standards. Summary of the Invention

[0004] This invention provides a dynamic intelligent evaluation system for the learning behavior of medical college students that integrates AI methods, solving the problems of evaluation lag and strong subjectivity in traditional medical education evaluation methods.

[0005] The basic solution provided by this invention is: a dynamic intelligent evaluation system for the learning behavior of medical college students that integrates AI methods, comprising: A multi-source data acquisition module is used to collect learning behavior data, which includes process-oriented learning behavior data and outcome-oriented learning performance data. The data preprocessing and fusion module is used to clean and standardize the learning behavior data, and to correlate the preprocessed data. The multimodal fusion analysis module, based on the processed learning behavior data, calls the multimodal fusion model to map the learning behavior data into dynamic quantitative scores of medical students' ability dimensions; The multimodal fusion model includes: The heterogeneous coding layer consists of multiple parallel encoders that extract procedural high-dimensional features and prior knowledge features from the learning behavior data, respectively. The collaborative fusion layer is used to learn the interaction relationships between procedural high-dimensional features through a cross-modal attention mechanism and dynamically fuse them to generate a procedural joint representation vector; The task adaptation layer, based on the process joint representation vector and prior knowledge features, outputs dynamic quantitative scores for multiple core competency dimensions through parallel regression of a multi-task learning network. The evaluation and feedback module is used to generate visualized dynamic profiles of individual and group abilities based on dynamic quantitative scores; it is also used to generate knowledge reinforcement suggestions based on dynamic quantitative scores and attention weights generated by the collaborative fusion layer; and it is used to identify individual student ability weaknesses based on learning behavior data to generate personalized learning path suggestions.

[0006] Preferably, the multi-source data acquisition module includes a practical data acquisition sub-module, a cognitive behavior data acquisition sub-module, and a learning performance data acquisition sub-module. The practical data acquisition submodule is used to collect physical operation behavior data in experimental operation scenarios and practical simulation scenarios. The physical operation behavior data includes operation video streams and / or sensor timing signals. The cognitive behavior data acquisition submodule is used to collect cognitive interaction behavior data in theoretical learning scenarios and simulated decision-making scenarios. The cognitive interaction behavior data includes online answer sequences, text interaction records, and structured decision-making path data. The learning performance data collection submodule is used to collect outcome-based learning performance data from the education information system. The outcome-based learning performance data includes academic performance data, learning input, and background data.

[0007] Preferably, in the data preprocessing and fusion module, the preprocessed data is correlated to form a multimodal data session dataset; The multimodal data session data generation strategy is as follows: 1) Unify the acquisition clock of the preprocessed learning behavior data to the same time base; 2) Analyze the feature changes of learning behavior data in parallel, determine the boundaries of candidate teaching events, and cut the learning behavior data into candidate event data segments according to the determined boundaries; 3) For each candidate event data segment, a dynamic time warping algorithm is used for time alignment, and the time-aligned candidate event data segment is input into a pre-trained event classification model to perform fusion feature recognition and event type labeling for each candidate event data segment; 4) Index and associate the aligned and labeled candidate event data segments with the metadata, and encapsulate the associated whole into a multimodal data session, and store it in the session database to form a multimodal data session dataset; the metadata includes event type labels, student unique identifiers, and absolute start and end timestamps of the event.

[0008] More preferably, in the heterogeneous coding layer, the encoder is configured according to the input data type. Specifically, procedural learning behavior data uses a visual encoder, a temporal encoder, and a semantic and sequence encoder for feature extraction, while outcome-based learning performance data uses a structured feature encoder for feature extraction. The visual encoder employs a spatiotemporal graph convolutional network to process operation video segments in structured multimodal data sessions, extracting high-dimensional features of operation posture and dynamics. The time-series encoder employs a one-dimensional dilated causal convolutional network to process sensor time-series signal segments in structured multimodal data sessions, extracting high-dimensional features of operational stability and rhythm. A semantic and sequence encoder, which is a composite structure containing: a) The text encoding submodule uses a pre-trained language model finely tuned on a medical corpus to process natural language text in structured multimodal data conversations and extract semantic features; b) The graph coding submodule uses a graph attention network to process decision path graph data in structured multimodal data sessions and extract decision logic features; c) Event sequence encoding submodule, which uses Transformer encoder to process online interaction log sequences in structured multimodal data sessions and extract learning behavior pattern features; The structured feature encoder uses a multilayer perceptron and a temporal modeling network to process preprocessed outcome learning performance data and extract a high-dimensional feature vector that comprehensively represents students' knowledge baseline, learning habits and long-term performance trends, which serves as the prior knowledge feature vector.

[0009] Preferably, the process joint representation vector fusion strategy is as follows: 1) The process feature vectors are mapped to a unified common feature space through independent learnable linear projection layers to obtain the projected feature vector set; 2) Calculate the query vector, key vector, and value vector of the projected features for each modality in the feature vector set, and calculate the gating scalar between the two modalities; 3) Calculate cross-modal attention weights; 4) Calculate the context vector of the modality based on the attention weights, fuse it with the original projection features to generate the enhanced representation of the modality, and aggregate all the enhanced representations to generate a global procedural joint representation vector.

[0010] More preferably, the cross-modal attention weight calculation process is as follows: First, calculate the initial attention score using the following formula;

[0011] In the formula, Let i be the query vector for modality i. Let j be the key vector of mode j. Scaling factor The dimension of the key vector; Then, normalization is performed using softmax combined with a gating scalar to obtain the attention weights of mode i for all modes. , specifically,

[0012] In the formula, Represents the set of all modalities. Indicates to Perform exponentiation. Let n be a gated scalar, and let n represent a modal index in set M.

[0013] Preferably, the input parameters of the multi-task learning network include a procedural joint representation vector and a prior knowledge feature vector. The multi-task learning network first fuses the procedural joint representation vector and the prior knowledge feature vector through a feature fusion unit to generate the final context-aware representation, and then outputs dynamic quantitative scores for each dimension through K independent prediction heads.

[0014] Preferably, the multi-task learning network is trained by minimizing a weighted total loss function, which is:

[0015] In the formula, K represents the number of core capability dimensions. The loss weight for the k-th task. Let be the loss function corresponding to the k-th task; for regression tasks, mean squared error loss or Huber loss is used; for classification tasks, cross-entropy loss is used.

[0016] Preferably, the evaluation and feedback module includes an interpretability feedback generation submodule, a real-time feedback triggering submodule, and a personalized path recommendation submodule; The interpretability feedback generation submodule is used to parse the attention weights generated by the collaborative fusion layer, associate the basis of the dynamic quantitative scoring with specific segments in the original learning behavior data, and generate natural language descriptions or visual highlighting prompts containing the basis. The real-time feedback triggering submodule is used to monitor the practical operation and simulated training process. When the dynamic quantitative score deviates from the preset threshold, it generates and pushes correction guidance information in real time. The personalized path recommendation submodule is used to generate learning resources and task sequences that are suitable for the student's current state based on historical learning behavior data and ability shortcomings, using reinforcement learning or sequence recommendation algorithms.

[0017] More preferably, the real-time feedback triggering submodule uses a contextual gambling machine algorithm to optimize the feedback strategy online. The contextual gambling machine algorithm uses the improvement of the student's subsequent ability score as a reward signal and dynamically adjusts the type and intensity of the triggered feedback in different situations.

[0018] The principles and advantages of this invention are as follows: 1. By continuously collecting and analyzing process data, the system can achieve real-time and continuous tracking and scoring of students' ability development, overcoming the lag of traditional summative assessment and providing the possibility for timely teaching intervention. Moreover, the data collection covers all scenarios of theoretical learning, experimental operation, and practical simulation, and by using multimodal fusion analysis technology, it realizes dynamic and continuous tracking and quantitative evaluation of students' knowledge, skills, and qualities, overcoming the limitations of traditional static and single-dimensional evaluation.

[0019] 2. A hierarchical multimodal fusion model is adopted to perform hierarchical fusion of process high-dimensional features and prior knowledge features, which automatically and objectively analyzes complex heterogeneous behavioral data, reduces the subjective dependence on manual evaluation, and significantly improves the consistency and scientificity of the evaluation.

[0020] 3. Based on dynamic quantitative scoring and interpretable attention weights, it can provide targeted real-time operational correction guidance and personalized learning path suggestions based on individual ability shortcomings, realizing a closed loop from evaluation to learning improvement, which greatly improves the timeliness and effectiveness of teaching intervention. Attached Figure Description

[0021] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0022] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: (See details) Figure 1 A dynamic intelligent evaluation system for the learning behavior of medical college students that integrates AI methods includes a multi-source data acquisition module, a data preprocessing and fusion module, a multimodal fusion analysis module, and an evaluation and feedback module.

[0023] The multi-source data acquisition module is used to collect multi-source heterogeneous learning behavior data. The learning behavior data includes process learning behavior data and outcome learning performance data. The process learning behavior data includes practical data and cognitive behavior data. The outcome learning performance data includes outcome learning performance data and outcome learning performance data.

[0024] The multi-source data acquisition module includes a practical data acquisition sub-module, a cognitive behavior data acquisition sub-module, and a learning performance data acquisition sub-module. The practical data acquisition submodule is used to collect physical operation behavior data in experimental operation scenarios and practical simulation scenarios. The physical operation behavior data includes operation video streams and / or sensor timing signals. Specifically, the physical operation actions of students in real or simulated environments are recorded by sensors and cameras. The high-definition video stream of the operation process, the timing signals from intelligent teaching aids (such as force feedback surgical instruments, simulated syringes, and data on force, displacement, angle, etc.), and the operation logs and status parameters of experimental equipment or highly realistic mannequins are used as practical data.

[0025] The cognitive behavior data acquisition submodule is used to collect cognitive interaction behavior data in theoretical learning scenarios and simulated decision-making scenarios. The cognitive interaction behavior data includes online answer sequences, text interaction records, and structured decision-making path data. Specifically, we acquire students' interactive behaviors on platforms such as the Learning Management System (LMS), Virtual Patient (VP) system, and online discussion forums. We use structured event logs, such as answer sequences (question ID, answer, time taken), video viewing behaviors (play, pause, jump), text data (such as case analysis reports, forum discussion content, and consultation record text), student choice path graphs (nodes are decision points, edges are choices) in the virtual patient or decision simulation system, and audio recordings of simulated consultations or team training (after anonymization) as cognitive behavioral data. The learning performance data collection submodule is used to collect outcome-based learning performance data from the education information system. The outcome-based learning performance data includes academic performance data, learning engagement and background data. Specifically, academic performance data is quantitative evaluation results obtained through interfaces from the academic affairs system, online examination system, and OSCE assessment system, including but not limited to test scores, assignment scores, lab report scores, stage assessment scores, and overall course scores. Learning engagement and background data are learning engagement indicators statistically collected from the learning management system, such as total login time, number of resource downloads, and forum participation. Student background information is also obtained from the student registration system, such as major, grade, courses taken, and their grades.

[0026] The data preprocessing and fusion module is used to clean and standardize the learning behavior data, and to correlate the preprocessed data.

[0027] The data preprocessing and fusion module includes a data cleaning and standardization submodule and a multimodal spatiotemporal alignment and correlation submodule. The data cleaning and standardization submodule is used to clean and standardize the learning behavior data. The data cleaning includes missing value handling and outlier detection and handling.

[0028] In this embodiment, the data type is first identified, and different types of data are processed differently according to a preset data processing strategy: For video stream data, key frame filtering based on time consistency and optical flow analysis are used for data cleaning. Specifically, the optical flow changes between video frames are calculated, and long periods of time with optical flow amplitude continuously below the threshold are regarded as invalid static or redundant segments and are removed. At the same time, violent shaking and mosaic frames caused by equipment jitter or transmission errors are detected and removed, and the video sequence that best reflects the operation process is retained. For sensor time-series signals, wavelet transform denoising and sliding window anomaly detection based on physiological operational constraints are used for data cleaning. Specifically, firstly, wavelet threshold denoising is used to filter out high-frequency electronic noise and power frequency interference; secondly, hard thresholds are set according to the ergonomic limits of medical operations (such as the maximum applicable force range and the range of joint movement angles) to remove outliers that clearly exceed physical possibilities; finally, the statistical characteristics of the signal (such as mean and variance) are calculated within the sliding window, and the isolated forest algorithm is used to detect and process segments with local abnormalities. For log / sequence data, a business logic-based integrity verification and sequence reconstruction method is used for cleaning. Specifically, for example, verifying the logical order and paired occurrence of the "start answering" and "submit answer" events in the answer sequence, correcting out-of-order timestamps caused by network latency; and mapping and aligning heterogeneous log fields collected from different platforms to a unified internal standard data model. For text data (such as case analysis reports and forum discussion content), a word segmentation and semantic error correction method combining a medical thesaurus is adopted. Specifically, a word segmentation tool loaded with a medical terminology dictionary is used to segment words to ensure accurate identification of professional entities; and a pre-trained medical language model is used to detect potential spelling errors and contextual semantic contradictions in the input text and provide suggested corrections. For audio data, an endpoint detection and noise reduction method based on acoustic features is adopted. Speech activity detection (VAD) technology is used to separate effective speech segments from silent / noise segments, and spectral subtraction is used to suppress environmental background noise, preserving clear speech signals for subsequent analysis. For structured tabular data (such as academic performance, learning engagement indicators, etc.), a data cleaning method based on statistical rules and domain knowledge is adopted. Specifically, this includes detecting and processing illegal values ​​in the performance data; for a missing performance, making a reasonable estimate or marking it by combining the student's other performance trends and course weights; and standardizing background information (such as major names) from different systems into a unified code. The standardization process involves Z-score standardization of all cleaned numerical features, converting them into a standard distribution with a mean of 0 and a standard deviation of 1, in order to eliminate the influence of dimensions.

[0029] The multimodal spatiotemporal alignment and association submodule is used to associate the preprocessed data to form a multimodal data session dataset; The multimodal data session data generation strategy is as follows: 1) By using the Network Time Protocol (NTP) or hardware synchronization signals, the acquisition clocks of the preprocessed learning behavior data are unified to the same time base. In this embodiment, the Network Time Protocol is used to synchronize the clocks of all acquisition devices (such as cameras, sensors, and simulation system servers) to a unified time source, so as to establish a common time base for cross-device alignment. 2) Parallel analysis of the feature changes in learning behavior data to determine the boundaries of candidate teaching events, and based on the determined boundaries, the learning behavior data is segmented into candidate event data segments; specifically, the temporal features representing the activity intensity in the learning behavior data are extracted respectively. When the feature values ​​of two or more modalities are detected to exceed their adaptive thresholds simultaneously within a similar time window, it is determined that a significant teaching state switch has occurred, and this time point is marked as a candidate event boundary point; based on two consecutive boundary points (start point and end point), the continuous data stream is segmented into independent candidate event data segments; 3) For each candidate event data segment, a dynamic time warping algorithm is used for time alignment. The time-aligned candidate event data segments are then input into a pre-trained event classification model. Feature fusion and event type labeling are performed on each candidate event data segment. Specifically, the dynamic time warping algorithm is used to perform frame-level or sampling-point-level fine time alignment on modal data with different sampling rates within the segment, eliminating minor misalignments caused by transmission delays or differences in acquisition frequency, and generating time-synchronized multimodal data segments. The time-synchronized multimodal data segments are then input into a pre-trained multimodal event classification model. This model fuses the synchronization features of each modality, identifies the semantic content of the data segment, and labels it with standardized event type tags (such as "skill practice_microinjection practice", "simulated diagnosis and treatment_acute chest pain management"). Finally, the output is a time-aligned multimodal data object with clear event type labels. 4) Index and associate the aligned and labeled candidate event data segments with the metadata, and encapsulate the associated whole into a multimodal data session, and store it in the session database to form a multimodal data session dataset. Each session is the smallest logical unit for the system to perform fusion analysis and dynamic evaluation. The metadata includes event type labels, student unique identifiers, and absolute start and end timestamps of the event.

[0030] The multimodal fusion analysis module, based on the processed learning behavior data, calls the multimodal fusion model to map the learning behavior data into dynamic quantitative scores of medical students' ability dimensions; The multimodal fusion model includes: A heterogeneous coding layer includes multiple parallel encoders, each extracting high-dimensional features from the learning behavior data; the encoders are configured according to the input data type, including: The visual encoder employs a spatiotemporal graph convolutional network to process operation video segments in structured multimodal data sessions, extracting high-dimensional features of operation posture and dynamics. The time-series encoder employs a one-dimensional dilated causal convolutional network to process sensor time-series signal segments in structured multimodal data sessions, extracting high-dimensional features of operational stability and rhythm. A semantic and sequence encoder, which is a composite structure containing: a) The text encoding submodule uses a pre-trained language model finely tuned on a medical corpus to process natural language text in structured multimodal data conversations and extract semantic features; b) The graph coding submodule uses a graph attention network to process decision path graph data in structured multimodal data sessions and extract decision logic features; c) Event sequence encoding submodule, which uses Transformer encoder to process online interaction log sequences in structured multimodal data sessions and extract learning behavior pattern features; The structured feature encoder employs a multilayer perceptron and a temporal modeling network to specifically process preprocessed outcome-based learning performance data (including academic performance data, learning input, and background data). It extracts a high-dimensional feature vector that comprehensively represents students' knowledge baseline, learning habits, and long-term performance trends, serving as a prior knowledge feature vector. Specifically, it extracts static contextual features by embedding and encoding learning input and background data using a multilayer perceptron, and extracts dynamic evolutionary features by encoding the temporal sequence composed of academic performance data using a recurrent neural network. Finally, it fuses the two to generate a high-dimensional feature vector that comprehensively represents students' knowledge baseline, learning habits, and long-term performance trends.

[0031] The collaborative fusion layer is used to learn the interaction relationships between procedural high-dimensional features through a cross-modal attention mechanism and dynamically fuse them to generate a procedural joint representation vector. These procedural high-dimensional features include output features from the visual encoder, temporal encoder, and semantic and sequence encoder (containing text, graph, and event sequence sub-modules), denoted as follows: ; The process joint representation vector fusion strategy is as follows: 1) Process feature vector Each feature is mapped to a unified common feature space through an independent learnable linear projection layer. The projected feature vector set is obtained. The projection formula is:

[0032] In the formula, , For the projection weight matrix and bias vector for mode m, These are learnable modality-specific projection parameters. It serves as the common feature space dimension; by projection, the differences in the original dimensions and distributions of different feature vectors are eliminated, providing a comparable basis for subsequent cross-modal interactions; 2) Calculate the query vector, key vector, and value vector of the projected features for each modality, and calculate the gating scalar between the two modalities; specifically,

[0033]

[0034]

[0035]

[0036] In the formula, For the gated scalar between mode i and mode j, , , It is a shared linearly changing weight matrix. The dimension of the key vector. This represents vector concatenation. and These are learnable parameters. It is the Sigmoid function, which compresses the output to the (0,1) interval; 3) Calculate cross-modal attention weights; Specifically, the initial attention score is first calculated using the following formula;

[0037] In the formula, Let i be the query vector for modality i. Let j be the key vector of mode j. Scaling factor The dimension of the key vector (i.e.); Then, normalization is performed using softmax combined with a gating scalar to obtain the attention weights of mode i for all modes. :

[0038] In the formula, Represents the set of all modalities. Indicates to Perform exponentiation. For gated scalars, n represents a modal index in set M; 4) Calculate the context vector of the modality based on the attention weights, fuse it with the original projected features to generate an enhanced representation of the modality, and aggregate all enhanced representations to generate a global procedural joint representation vector; Specifically, the context vector for each modality The value vectors of all modalities are obtained by weighted summation based on the attention weights:

[0039] In the formula, Subsequently, the original projection features are fused with their contextual information to generate an enhanced representation of the modality. :

[0040] Here, FFN is a two-layer feedforward neural network, and LayerNorm is a layer normalization operation.

[0041] Finally, the enhanced representations of all modalities are aggregated to generate a global procedural joint representation vector. This embodiment uses additive pooling:

[0042] In the formula, This is the global joint representation vector; Process joint representation vector It is sent to the task adaptation layer for final capability dimension score prediction.

[0043] The task adaptation layer, based on procedural joint representation vectors and prior knowledge features, outputs dynamic quantitative scores for multiple core competency dimensions through parallel regression of a multi-task learning network.

[0044] The input to the task adaptation layer includes two parts: 1) the procedural joint representation vector from the collaborative fusion layer. This vector integrates the multimodal procedural features of the learning event; 2) the prior knowledge feature vector from the output of the structured feature encoder. This vector encodes the student's historical academic performance, learning engagement, and background information; The multi-task learning network first uses a feature fusion unit to combine the procedural joint representation vector. With prior knowledge feature vector The features are fused to generate a final context-aware representation, which is then output as dynamic quantitative scores for each dimension via K independent prediction heads. The feature fusion unit can employ strategies such as concatenating the features and then connecting them to a fully connected layer, or gating and modulating the procedural representation based on prior knowledge.

[0045] The multi-task learning network is trained by minimizing a weighted total loss function, which is:

[0046] In the formula, K represents the number of core capability dimensions. The loss weight for the k-th task. Let be the loss function corresponding to the k-th task; for regression tasks, mean squared error loss or Huber loss is used; for classification tasks, cross-entropy loss is used.

[0047] The evaluation and feedback module is used to generate visualized dynamic profiles of individual and group abilities based on dynamic quantitative scores; it is also used to generate personalized cognitive reinforcement suggestions based on dynamic quantitative scores and attention weights generated by the collaborative fusion layer; and it is used to identify individual student ability weaknesses based on learning behavior data to generate personalized learning path suggestions.

[0048] The evaluation and feedback module includes an interpretable feedback generation submodule, a real-time feedback triggering submodule, and a personalized path recommendation submodule; The interpretability feedback generation submodule is used to parse the attention weights generated by the collaborative fusion layer, associate the basis of the dynamic quantitative scoring with specific segments in the original learning behavior data, and generate natural language descriptions or visual highlighting prompts containing the basis. Specifically, after a "laparoscopic cholecystectomy simulation training," Student A received a score of 55 / 100 in the "tissue separation precision" dimension. Simultaneously, the system retrieved the student's historical grades, showing excellent theoretical performance in "Fundamentals of Surgery" but an average score of only 60 in "Surgical Procedures." The fusion layer attention weights were also retrieved. = 0.50, = 0.35, = 0.15, indicating that the video modality (operational movements) was the main reason for the low score. Further tracing the source and generating feedback information, it was found that: "Your score of 'Tissue Separation Fineness' (55 points) is significantly lower than your theoretical level (90 points in 'Fundamentals of Surgery'), which may reflect the difficulty in 'converting knowledge into practice'; the main operational problem is: in the middle of the operation (10:05-10:07), your operating instruments showed unnecessary lateral oscillation (see highlighted video), which increased the risk of tissue damage; Recommendation: Focus on practicing the transformation of theoretical knowledge into stable operation, and conduct small-amplitude, high-frequency 'push-open' movement training." This feedback explains the low score in the context of the individual's "strong theory, weak operation", which is more personalized guidance.

[0049] The real-time feedback triggering submodule is used to monitor the practical operation and simulated training process. When the dynamic quantitative score deviates from the preset threshold, it generates and pushes correction guidance information in real time. The preset thresholds are pre-set based on domain knowledge (expert experience) or historical data statistics. For example, the threshold for hand tremor frequency is 5 Hz (tremors exceeding this frequency will affect fine motor skills); the threshold for suture stitch distance variance is 2 mm (from surgical teaching standards); and the threshold for missing key decision-making steps is set so that if the "inquire about allergy history" step is missing, an early warning will be triggered directly (from the core medical safety system). Specifically, Student B is undergoing a simulated training session on "central venous catheterization." The real-time video stream shows that the puncture needle has entered the preset anatomical area, and the real-time pressure sensor signal shows a sudden drop in needle tip resistance (potentially indicating puncture into a vein). However, the real-time video analysis model detects that Student B's hand holding the needle shows a tendency to continue advancing (a key behavioral characteristic). The system compares the real-time sensor characteristic (sudden drop in resistance) with a standard feature library and determines that "the needle has successfully entered the blood vessel." At the same time, it detects the key behavioral characteristic that "the hand still has a preparatory movement to advance forward," which deviates from the safe operating threshold of "stopping needle advancement immediately after entering the blood vessel." The system generates corrective guidance information (triggered within 0.5 seconds before the action occurs): "Real-time warning: Detection of blood vessel wall breakthrough! Please stop advancing the needle immediately! Current action: Sensor shows that the resistance has decreased, indicating that the needle tip has entered the blood vessel lumen; Risk: Continuing to advance the needle may lead to posterior wall penetration of the blood vessel; Corrective action: Fix the needle body and prepare to insert the guidewire."

[0050] The real-time feedback triggering submodule uses a contextual gambling machine algorithm to optimize the feedback strategy online. The contextual gambling machine algorithm uses the improvement of the student's subsequent ability score as a reward signal and dynamically adjusts the type and intensity of the triggered feedback in different situations. The personalized path recommendation submodule is used to generate learning resources and task sequences that are suitable for the student's current state based on historical learning behavior data and ability shortcomings, using reinforcement learning or sequence recommendation algorithms.

[0051] The system also includes: The data storage and management module is used to store structured multimodal data sessions, dynamic quantitative scoring sequences, and evaluation feedback records in a graph structure or relational database, and supports multi-dimensional query and analysis based on teaching events, students, and ability dimensions.

[0052] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods, characterized in that, include: A multi-source data acquisition module is used to collect learning behavior data, which includes process-oriented learning behavior data and outcome-oriented learning performance data. The data preprocessing and fusion module is used to clean and standardize the learning behavior data, and to correlate the preprocessed data. The multimodal fusion analysis module, based on the processed learning behavior data, calls the multimodal fusion model to map the learning behavior data into dynamic quantitative scores of medical students' ability dimensions; The multimodal fusion model includes: The heterogeneous coding layer consists of multiple parallel encoders that extract procedural high-dimensional features and prior knowledge features from the learning behavior data, respectively. The collaborative fusion layer is used to learn the interaction relationships between procedural high-dimensional features through a cross-modal attention mechanism and dynamically fuse them to generate a procedural joint representation vector; The task adaptation layer, based on the process joint representation vector and prior knowledge features, outputs dynamic quantitative scores for multiple core competency dimensions through parallel regression of a multi-task learning network. The evaluation and feedback module is used to generate visualized dynamic profiles of individual and group abilities based on dynamic quantitative scores; it is also used to generate knowledge reinforcement suggestions based on dynamic quantitative scores and attention weights generated by the collaborative fusion layer; and it is used to identify individual student ability weaknesses based on learning behavior data to generate personalized learning path suggestions.

2. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 1, characterized in that: The multi-source data acquisition module includes a practical data acquisition sub-module, a cognitive behavior data acquisition sub-module, and a learning performance data acquisition sub-module. The practical data acquisition submodule is used to collect physical operation behavior data in experimental operation scenarios and practical simulation scenarios. The physical operation behavior data includes operation video streams and / or sensor timing signals. The cognitive behavior data acquisition submodule is used to collect cognitive interaction behavior data in theoretical learning scenarios and simulated decision-making scenarios. The cognitive interaction behavior data includes online answer sequences, text interaction records, and structured decision-making path data. The learning performance data collection submodule is used to collect outcome-based learning performance data from the education information system. The outcome-based learning performance data includes academic performance data, learning input, and background data.

3. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 1, characterized in that: In the data preprocessing and fusion module, the preprocessed data is correlated to form a multimodal data session dataset; The multimodal data session data generation strategy is as follows: 1) Unify the acquisition clock of the preprocessed learning behavior data to the same time base; 2) Analyze the feature changes of learning behavior data in parallel, determine the boundaries of candidate teaching events, and cut the learning behavior data into candidate event data segments according to the determined boundaries; 3) For each candidate event data segment, a dynamic time warping algorithm is used for time alignment, and the time-aligned candidate event data segment is input into a pre-trained event classification model to perform fusion feature recognition and event type labeling for each candidate event data segment; 4) Index and associate the aligned and labeled candidate event data segments with the metadata, and encapsulate the associated whole into a multimodal data session, and store it in the session database to form a multimodal data session dataset; the metadata includes event type labels, student unique identifiers, and absolute start and end timestamps of the event.

4. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 2, characterized in that: In the heterogeneous coding layer, encoders are configured according to the input data type. Specifically, procedural learning behavior data uses a visual encoder, a temporal encoder, and a semantic and sequence encoder for feature extraction, while outcome learning performance data uses a structured feature encoder for feature extraction. The visual encoder employs a spatiotemporal graph convolutional network to process operation video segments in structured multimodal data sessions, extracting high-dimensional features of operation posture and dynamics. The time-series encoder employs a one-dimensional dilated causal convolutional network to process sensor time-series signal segments in structured multimodal data sessions, extracting high-dimensional features of operational stability and rhythm. A semantic and sequence encoder, which is a composite structure containing: a) The text encoding submodule uses a pre-trained language model finely tuned on a medical corpus to process natural language text in structured multimodal data conversations and extract semantic features; b) The graph coding submodule uses a graph attention network to process decision path graph data in structured multimodal data sessions and extract decision logic features; c) Event sequence encoding submodule, which uses Transformer encoder to process online interaction log sequences in structured multimodal data sessions and extract learning behavior pattern features; The structured feature encoder uses a multilayer perceptron and a temporal modeling network to process preprocessed outcome learning performance data and extract a high-dimensional feature vector that comprehensively represents students' knowledge baseline, learning habits and long-term performance trends, which serves as the prior knowledge feature vector.

5. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 1, characterized in that: The process joint representation vector fusion strategy is as follows: 1) The process feature vectors are mapped to a unified common feature space through independent learnable linear projection layers to obtain the projected feature vector set; 2) Calculate the query vector, key vector, and value vector of the projected features for each modality in the feature vector set, and calculate the gating scalar between the two modalities; 3) Calculate cross-modal attention weights; 4) Calculate the context vector of the modality based on the attention weights, fuse it with the original projection features to generate the enhanced representation of the modality, and aggregate all the enhanced representations to generate a global procedural joint representation vector.

6. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 5, characterized in that: The cross-modal attention weight calculation process is as follows: First, calculate the initial attention score using the following formula; In the formula, Let i be the query vector for modality i. Let j be the key vector of mode j. Scaling factor The dimension of the key vector; Then, normalization is performed using softmax combined with a gating scalar to obtain the attention weights of mode i for all modes. , specifically, In the formula, Represents the set of all modalities. Indicates to Perform exponentiation. Let n be a gated scalar, and let n represent a modal index in set M.

7. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 1, characterized in that: The input parameters of the multi-task learning network include the procedural joint representation vector and the prior knowledge feature vector. The multi-task learning network first fuses the procedural joint representation vector and the prior knowledge feature vector through a feature fusion unit to generate the final context-aware representation, and then outputs the dynamic quantitative scores of each dimension through K independent prediction heads.

8. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 1, characterized in that: The multi-task learning network is trained by minimizing a weighted total loss function, which is: In the formula, K represents the number of core capability dimensions. The loss weight for the k-th task. Let be the loss function corresponding to the k-th task; for regression tasks, mean squared error loss or Huber loss is used; for classification tasks, cross-entropy loss is used.

9. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 1, characterized in that: The evaluation and feedback module includes an interpretable feedback generation submodule, a real-time feedback triggering submodule, and a personalized path recommendation submodule; The interpretability feedback generation submodule is used to parse the attention weights generated by the collaborative fusion layer, associate the basis of the dynamic quantitative scoring with specific segments in the original learning behavior data, and generate natural language descriptions or visual highlighting prompts containing the basis. The real-time feedback triggering submodule is used to monitor the practical operation and simulated training process. When the dynamic quantitative score deviates from the preset threshold, it generates and pushes correction guidance information in real time. The personalized path recommendation submodule is used to generate learning resources and task sequences that are suitable for the student's current state based on historical learning behavior data and ability shortcomings, using reinforcement learning or sequence recommendation algorithms.

10. The dynamic intelligent evaluation system for the learning behavior of medical college students integrating AI methods according to claim 9, characterized in that: The real-time feedback triggering submodule uses a contextual gambling machine algorithm to optimize the feedback strategy online. The contextual gambling machine algorithm uses the improvement of the student's subsequent ability score as a reward signal and dynamically adjusts the type and intensity of the triggered feedback in different situations.

Images