Medical student tiredness risk early warning and intervention system based on dynamic data monitoring

Abstract

The invention discloses a medical student tiredness risk early warning and intervention system based on dynamic data monitoring, and the system comprises an obtaining module which is used for obtaining the dynamic monitoring data and static basic data of a to-be-evaluated medical student; the preprocessing module is used for preprocessing the dynamic monitoring data to obtain preprocessed dynamic monitoring data; the prediction module is used for inputting the static basic data and the preprocessed dynamic monitoring data into a pre-trained tiredness risk prediction model for prediction, and outputting a tiredness risk level of the to-be-evaluated medical student; the early warning module is used for judging whether an early warning prompt is sent out or not based on the tiredness risk level of the to-be-evaluated medical student; the intervention module is used for obtaining an intervention strategy corresponding to the tiredness risk level of the to-be-assessed medical student when it is determined that the early warning prompt is sent out, and intervening the to-be-assessed medical student based on the intervention strategy; the fatigue risk of the medical students is analyzed through multi-dimensional information, early warning is conducted in time when the medical students have slight fatigue symptoms, and the intervention effect is improved.

Description

Technical Field

[0001] This invention relates to the field of burnout risk prediction and intervention technology, and in particular to a burnout risk early warning and intervention system for medical students based on dynamic data monitoring. Background Technology

[0002] Medical students, as the core reserve force of the future healthcare system, must simultaneously bear the heavy burden of high-intensity academic studies (such as complex medical theory learning and experimental operation training) and clinical practice pressure (such as night shifts, patient communication, and emergency handling), placing them under a long-term dual load of "academics-clinical" work. In recent years, numerous studies both domestically and internationally have shown that the incidence of burnout among medical students is significantly higher than that of the general university student population. Burnout symptoms are mainly manifested as emotional exhaustion, depersonalization (indifference towards patients or classmates), and decreased personal achievement. This not only directly harms their physical and mental health (such as increased risk of anxiety and depression, and sleep disorders), but also leads to decreased academic efficiency (increased failure rates and increased practical operation errors), and may even affect the quality of future medical practice (such as insufficient patience in doctor-patient communication and increased risk of medical errors), posing a potential threat to the training of medical and health personnel and the development of the industry. Therefore, establishing an effective mechanism for monitoring and early intervention of medical student burnout risks has become a key need in the field of medical education management.

[0003] First, current technologies for assessing burnout risk in medical students can only be used to evaluate and intervene when medical students have already shown obvious burnout symptoms. They cannot provide timely warnings when medical students have mild burnout symptoms, which leads to missing the best time for intervention and resulting in poor intervention effects. Second, current technologies for assessing burnout risk have relatively limited information dimensions, leading to inaccurate warning results.

[0004] Therefore, there is an urgent need for a medical student burnout risk early warning and intervention system based on dynamic data monitoring to solve the above problems. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. Therefore, the purpose of this invention is to propose a medical student burnout risk early warning and intervention system based on dynamic data monitoring. This system analyzes the burnout risk of medical students from multiple dimensions, enabling timely early warning when medical students exhibit mild burnout symptoms, thereby improving the effectiveness of intervention.

[0006] To achieve the above objectives, embodiments of the present invention propose a medical student burnout risk early warning and intervention system based on dynamic data monitoring, comprising: The acquisition module is used to acquire dynamic monitoring data and static basic data of the medical students to be evaluated. The preprocessing module is used to preprocess the dynamic monitoring data to obtain preprocessed dynamic monitoring data; The prediction module is used to input static basic data and preprocessed dynamic monitoring data into a pre-trained burnout risk prediction model for prediction, and output the burnout risk level of the medical student to be evaluated. The early warning module is used to determine whether to issue an early warning based on the burnout risk level of the medical student to be assessed. The intervention module is used to obtain the intervention strategy corresponding to the burnout risk level of the medical student to be evaluated when it is determined to issue an early warning prompt, and to intervene on the medical student to be evaluated based on the intervention strategy.

[0007] Preferably, the acquisition module includes: The first acquisition submodule is used to acquire dynamic behavioral data of the medical students to be evaluated; the dynamic behavioral data includes campus card usage record data, library behavior data, and online learning platform data. The second acquisition submodule is used to acquire physiological status data of the medical student to be evaluated; the physiological status data includes sleep structure data, activity level data, and heart rate variability data. The determination submodule is used to use the dynamic behavioral data and physiological status data of the medical students to be evaluated as dynamic monitoring data of the medical students to be evaluated. The third acquisition submodule is used to acquire the static basic data of the medical students to be evaluated; the static basic data includes the grade, major, and previous psychological counseling center records of the medical students to be evaluated.

[0008] Preferably, the preprocessing module includes: The data cleaning submodule is used to clean the dynamic monitoring data to obtain cleaned dynamic monitoring data. The noise reduction submodule is used to reduce the noise of the cleaned dynamic monitoring data to obtain processed dynamic monitoring data.

[0009] Preferred training methods for burnout risk prediction models include: Obtain the training dataset for the burnout risk prediction model; Feature extraction is performed on the training dataset to obtain the multimodal training features corresponding to the training dataset; The multimodal training features are input into the multi-source data feature encoder of the dynamic hierarchical multi-task attention fusion model for feature encoding, and the multimodal training features are converted into multimodal feature vectors. The cross-modal attention fusion layer based on the dynamic hierarchical multi-task attention fusion model fuses and reduces the dimensionality of multimodal feature vectors to obtain cross-modal fused feature vectors. The cross-modal fusion feature vector is input into the multi-task collaborative prediction layer of the dynamic hierarchical multi-task attention fusion model to predict the burnout risk level, and the burnout risk level prediction result is output. The training dataset of the burnout risk prediction model is sequentially input into the dynamic hierarchical multi-task attention fusion model for training. When the training results meet the requirements, the initial burnout risk prediction model is obtained. Obtain the test dataset for the burnout risk prediction model; The initial burnout risk prediction model was tested using a test dataset based on the burnout risk prediction model. When the test results met the requirements, the trained burnout risk prediction model was obtained.

[0010] Preferably, the multimodal training features are input into the multi-source data feature encoder of the dynamic hierarchical multi-task attention fusion model for feature encoding, converting the multimodal training features into multimodal feature vectors, including: Based on the time-series physiological data encoder, the physiological state data included in the multimodal training features are feature-encoded to generate physiological feature vectors; Based on the discrete behavior graph encoder, feature encoding is performed on the campus card and library dynamic behavior feature data included in the multimodal training features to generate behavior feature vectors; Based on the learning activity Transformer encoder, feature encoding is performed on the online learning platform feature data included in the multimodal training features to generate learning feature vectors; Based on the static feature encoder, the grade, major and previous psychological counseling center records included in the multimodal training features are feature encoded to generate static feature vectors; Multimodal feature vectors are constructed based on physiological feature vectors, behavioral feature vectors, learning feature vectors, and static feature vectors.

[0011] Preferably, the cross-modal attention fusion layer based on the dynamic hierarchical multi-task attention fusion model fuses and reduces the dimensionality of multimodal feature vectors to obtain cross-modal fused feature vectors, including: The behavioral feature vector, the learning feature vector, and the static feature vector are mapped to the same dimension as the physiological feature vector, and a feature matrix is ​​constructed with the physiological feature vector. Construct a grade-level guidance attention bias matrix; For each cross-attention head, generate a query matrix, a key matrix, and a value matrix; calculate the cross-attention score based on the query matrix and key matrix; normalize each cross-attention score using the Softmax function to obtain the attention weight corresponding to each cross-attention head; output the fusion feature of each cross-attention head based on the value matrix and attention weight of each cross-attention head. The fusion features of several cross-attention heads are concatenated into a multi-dimensional matrix, and the cross-modal fusion feature vector is obtained by dimensionality reduction through convolutional layers.

[0012] Preferably, the cross-modal fusion feature vector is input into the multi-task collaborative prediction layer of the dynamic hierarchical multi-task attention fusion model to predict the burnout risk level, and the burnout risk level prediction result is output, including: The cross-modal fusion feature vector is input into the MLP network to obtain intermediate features. The intermediate features are divided into four sub-vectors: physiological factors, behavioral factors, learning factors, and grade factors. A risk factor attention pooling layer is constructed, and the attention weight of each sub-vector is calculated. The target intermediate vector is obtained by weighted summation based on the attention weights of each sub-vector. The target intermediate vector is input into the Softmax classification layer to obtain the burnout risk probability distribution. The risk level corresponding to the maximum burnout risk is used as the initial burnout prediction level, and the first loss function is calculated. Physiologically relevant sub-vectors are extracted from the cross-modal fusion feature vectors, input into the MLP network, outputting the sleep quality coefficient, and calculating the second loss function; Learn relevant sub-vectors are extracted from cross-modal fusion feature vectors, input into a 2-layer MLP network, output learning engagement coefficients, and a third loss function is calculated. Behavioral-physiological cross vectors are extracted from cross-modal fusion feature vectors, input into a 2-layer MLP network, output stress state coefficients, and a fourth loss function is calculated. Sleep quality coefficient, learning engagement coefficient, and stress state coefficient were used as auxiliary indicators. The target burnout risk level prediction results are determined based on the initial burnout prediction level, sleep quality coefficient, learning engagement coefficient, and stress state coefficient. A total loss function is constructed based on the first loss function, the second loss function, the third loss function, and the fourth loss function.

[0013] Preferably, the early warning module includes: The comparison submodule is used to compare the burnout risk level of the medical student to be evaluated with the preset warning threshold; The early warning submodule is used to issue an early warning when it is determined that the burnout risk level of the medical student to be evaluated is greater than the preset early warning threshold.

[0014] Preferably, the intervention module includes: The query submodule is used to query the risk level-intervention strategy table based on the burnout risk level of the medical student to be evaluated, and to determine the target intervention strategy. The intervention submodule is used to intervene in medical students under assessment based on goal-oriented intervention strategies.

[0015] Preferably, it also includes: an intervention assessment module, used to obtain the change index of the burnout risk level of the medical students within the preset time period after the intervention strategy is implemented on the medical students, and to dynamically adjust the intervention strategy for the medical students based on the change index.

[0016] This invention provides a medical student burnout risk early warning and intervention system based on dynamic data monitoring. By acquiring dynamic monitoring data and static basic data, and using predictive models for analysis, it can detect potential burnout risks in time before medical students show obvious burnout symptoms; predict the burnout risk level in advance, allowing medical students to receive attention and intervention in the early stages of burnout, preventing the problem from worsening and protecting their physical and mental health; formulate corresponding intervention strategies according to different burnout risk levels; each medical student's situation is unique, and the system can provide personalized intervention measures for each individual's specific risk level; it continuously collects and analyzes relevant data from medical students, which helps them better understand their learning status, psychological state, and behavioral patterns; through timely early warning and effective intervention, it reduces the risk of burnout for medical students, thereby improving their learning outcomes and professional competence.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of a medical student burnout risk early warning and intervention system based on dynamic data monitoring according to an embodiment of the present invention; Figure 2 This is a block diagram of an acquisition module according to an embodiment of the present invention; Figure 3 This is a block diagram of a preprocessing module according to an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Example 1: As Figure 1 As shown, a medical student burnout risk early warning and intervention system based on dynamic data monitoring includes: The acquisition module is used to acquire dynamic monitoring data and static basic data of the medical students to be evaluated. The preprocessing module is used to preprocess the dynamic monitoring data to obtain preprocessed dynamic monitoring data; The prediction module is used to input static basic data and preprocessed dynamic monitoring data into a pre-trained burnout risk prediction model for prediction, and output the burnout risk level of the medical student to be evaluated. The early warning module is used to determine whether to issue an early warning based on the burnout risk level of the medical student to be assessed. The intervention module is used to obtain the intervention strategy corresponding to the burnout risk level of the medical student to be evaluated when it is determined to issue an early warning prompt, and to intervene on the medical student to be evaluated based on the intervention strategy.

[0022] In this embodiment, dynamic monitoring data includes dynamic behavioral data and physiological state data; static basic data includes the grade, major, and previous records of the medical student to be evaluated at the psychological counseling center.

[0023] In this embodiment, preprocessing includes data noise reduction and data cleaning.

[0024] The working principle and beneficial effects of the above technical solution are as follows: By acquiring dynamic monitoring data and static basic data, and using predictive models for analysis, potential burnout risks can be detected in time before medical students show obvious symptoms of burnout; the burnout risk level can be predicted in advance, allowing medical students to receive attention and intervention in the early stages of burnout, preventing the problem from worsening and protecting their physical and mental health; corresponding intervention strategies can be formulated according to different burnout risk levels; each medical student's situation is unique, and the system can provide personalized intervention measures for each individual's specific risk level; relevant data from medical students will be continuously collected and analyzed, which helps medical students better understand their learning status, psychological state, and behavioral patterns; through timely warnings and effective interventions, the risk of burnout among medical students can be reduced, thereby improving their learning outcomes and professional competence.

[0025] Example 2: Figure 2 As shown, the acquisition module includes: The first acquisition submodule is used to acquire dynamic behavioral data of the medical students to be evaluated; the dynamic behavioral data includes campus card usage record data, library behavior data, and online learning platform data. The second acquisition submodule is used to acquire physiological status data of the medical student to be evaluated; the physiological status data includes sleep structure data, activity level data, and heart rate variability data. The determination submodule is used to use the dynamic behavioral data and physiological status data of the medical students to be evaluated as dynamic monitoring data of the medical students to be evaluated. The third acquisition submodule is used to acquire the static basic data of the medical students to be evaluated; the static basic data includes the grade, major, and previous psychological counseling center records of the medical students to be evaluated.

[0026] The working principle and beneficial effects of the above technical solution are as follows: it reflects the daily learning and living conditions of medical students from different perspectives. Campus card usage records can reflect students' consumption habits, dining patterns, and activity range on campus; library behavior data can show students' self-study situation, such as the types of books borrowed and the time spent in the library; online learning platform data can reflect students' online learning activity, course participation, and learning progress; the comprehensive acquisition of these data provides rich information for a comprehensive understanding of medical students' learning behavior patterns; the acquired physiological state data, such as sleep structure data, activity level data, and heart rate variability data, are directly related to medical students' physical health and psychological stress; poor sleep quality and insufficient activity levels... Abnormal heart rate variability can be a physiological manifestation of burnout. Combining dynamic behavioral data can more accurately depict the state of medical students from both physiological and behavioral perspectives, thereby improving the accuracy of burnout risk assessment. Static baseline data, such as grade, major, and previous records from the psychological counseling center, provides background information for the assessment. Medical students of different grades and majors face different learning pressures and course difficulties, and previous psychological counseling records can reflect students' past psychological conditions and potential problems. Combining these static data with dynamic monitoring data helps to construct a more comprehensive and accurate assessment of burnout risk for medical students.

[0027] Example 3: As Figure 3 As shown, the preprocessing module includes: The data cleaning submodule is used to clean the dynamic monitoring data to obtain cleaned dynamic monitoring data. The noise reduction submodule is used to reduce the noise of the cleaned dynamic monitoring data to obtain processed dynamic monitoring data.

[0028] The working principle and beneficial effects of the above technical solution are as follows: Through data cleaning and noise reduction, the quality and usability of dynamic monitoring data are improved, providing strong support for subsequent data analysis, modeling and decision-making, and helping to improve the performance and efficiency of the entire system.

[0029] Example 4: Training method for burnout risk prediction model, including: Obtain the training dataset for the burnout risk prediction model; Feature extraction is performed on the training dataset to obtain the multimodal training features corresponding to the training dataset; The multimodal training features are input into the multi-source data feature encoder of the dynamic hierarchical multi-task attention fusion model for feature encoding, and the multimodal training features are converted into multimodal feature vectors. The cross-modal attention fusion layer based on the dynamic hierarchical multi-task attention fusion model fuses and reduces the dimensionality of multimodal feature vectors to obtain cross-modal fused feature vectors. The cross-modal fusion feature vector is input into the multi-task collaborative prediction layer of the dynamic hierarchical multi-task attention fusion model to predict the burnout risk level, and the burnout risk level prediction result is output. The training dataset of the burnout risk prediction model is sequentially input into the dynamic hierarchical multi-task attention fusion model for training. When the training results meet the requirements, the initial burnout risk prediction model is obtained. Obtain the test dataset for the burnout risk prediction model; The initial burnout risk prediction model was tested using a test dataset based on the burnout risk prediction model. When the test results met the requirements, the trained burnout risk prediction model was obtained.

[0030] In this embodiment, the dynamic hierarchical multi-task attention fusion model includes a multi-source data feature encoder, a cross-modal attention fusion layer, and a multi-task collaborative prediction layer.

[0031] The working principle and beneficial effects of the above technical solution are as follows: Multimodal training features are obtained by extracting features from the training dataset, comprehensively considering various types of data features. This enables the model to capture information related to burnout risk from multiple perspectives, providing a more comprehensive and in-depth understanding than single-modal features, thereby improving the model's understanding and prediction capabilities regarding burnout risk. Converting multimodal training features into multimodal feature vectors achieves a unified representation of different modalities at the feature level. This facilitates subsequent fusion processing of data from different sources, breaking down barriers between different modalities and allowing the model to comprehensively utilize the advantages of multi-source data. Attention weights are dynamically allocated based on the importance of different modal features, and the multimodal feature vectors are fused and dimensionality reduced. Simultaneously, multiple related tasks are considered, such as predicting different levels of burnout risk. This multi-task learning approach allows the model to share information between different tasks during training, improving the model's generalization ability and prediction accuracy.

[0032] Example 5: The multimodal training features are input into the multi-source data feature encoder of the dynamic hierarchical multi-task attention fusion model for feature encoding, converting the multimodal training features into multimodal feature vectors, including: Based on the time-series physiological data encoder, the physiological state data included in the multimodal training features are feature-encoded to generate physiological feature vectors; Based on the discrete behavior graph encoder, feature encoding is performed on the campus card and library dynamic behavior feature data included in the multimodal training features to generate behavior feature vectors; Based on the learning activity Transformer encoder, feature encoding is performed on the online learning platform feature data included in the multimodal training features to generate learning feature vectors; Based on the static feature encoder, the grade, major and previous psychological counseling center records included in the multimodal training features are feature encoded to generate static feature vectors; Multimodal feature vectors are constructed based on physiological feature vectors, behavioral feature vectors, learning feature vectors, and static feature vectors.

[0033] In this embodiment, a physiological feature vector is generated by encoding the physiological state data included in the multimodal training features based on a time-series physiological data encoder. This includes: inputting physiological state features (sleep structure, HRV, activity level) as a 7-day sliding window; reconstructing the physiological data into a 5×168-dimensional time-series matrix at 5-minute time steps; adding a 5-dimensional grade bias vector generated by grade static feature mapping; inputting an improved Bi-LSTM network (hidden layer dimension 256, dropout probability = 0.2, activation function tanh) to capture short-term temporal dependencies of physiological data (such as the trend of a sharp drop in HRV 1 hour before sleep); inputting the 256×168-dimensional feature matrix output by Bi-LSTM into a 6-layer TransformerEncoder (multi-head attention head number = 8, feed-forward network dimension = 512, activation function GELU) to capture long-term correlations of physiological data (such as the cumulative effect of deep sleep duration < 2 hours for 3 consecutive days); and outputting a 512-dimensional physiological feature vector after global average pooling (GAP) processing.

[0034] In this embodiment, a discrete behavior graph encoder is used to encode the campus card and library dynamic behavior feature data included in the multimodal training features to generate a behavior feature vector. The input features are campus card and library dynamic behavior data with a 7-day sliding window. A temporal behavior graph (TBG) is constructed: nodes are defined as 12 behavior types (dining in the cafeteria = 1, studying in the library = 2, attending classes in the teaching building = 3, hospital internship = 4, etc.), the edge weight is the reciprocal of the time interval between two behaviors (the shorter the interval, the higher the weight), and the node attribute is the total stay time of the behavior over 7 days. A grade adjustment coefficient is added to each node attribute of the TBG (e.g., the grade adjustment coefficient for the hospital internship node is 1.5 for fifth-year students and 0.3 for first-year students). A 2-layer graph attention network (GAT) is input (hidden layer dimension = 128, number of attention heads = 4, dropout probability = 0.2), and the attention weights between nodes are calculated (e.g., the weight of "hospital internship → dining in the cafeteria" reflects the dining pattern after the internship). All node features are weighted and summed according to the attention weights, and a 256-dimensional behavior feature vector is output.

[0035] In this embodiment, the online learning platform feature data included in the multimodal training features is feature-encoded based on the learning activity Transformer encoder to generate a learning feature vector. This includes: inputting online learning platform data with a 7-day sliding window as input (splittered by course type), embedding basic medical courses, clinical medical courses, and elective courses separately (basic medical courses embedding dimension = 64, clinical medical courses = 128, elective courses = 64), and adding grade-course interaction bias (e.g., the grade interaction bias for clinical medical courses is 1.2 for senior students and 0.5 for freshmen); calculating course type attention weights (weight = grade coefficient × 0.6 + course access duration coefficient × 0.4), and performing weighted fusion on the embedded data; inputting a 4-layer TransformerEncoder (multi-head attention head count = 6, feed-forward network dimension = 256, dropout probability = 0.2) to capture the correlation between learning behavior and time (e.g., the surge in access to clinical medical courses 3 days before the exam); and outputting a 256-dimensional learning feature vector after temporal max pooling processing.

[0036] In this embodiment, a static feature encoder is used to encode the grade, major, and previous psychological counseling center records included in the multimodal training features to generate a static feature vector. The input features are: grade one-hot encoding (5-dimensional for undergraduate / 7-dimensional for undergraduate and graduate students), major embedding vector (32-dimensional), and previous psychological counseling records (0 / 1). The grade one-hot encoding is mapped to a 64-dimensional grade feature vector through an embedding layer, and concatenated with the 32-dimensional major embedding vector to form a 96-dimensional vector. A 3-layer MLP network is input (first layer dimension = 128, second layer = 64, third layer = 32, activation function is GELU, residual connections are added to each layer to avoid gradient vanishing). The output is a 32-dimensional grade static feature vector.

[0037] The working principle and beneficial effects of the above technical solution are as follows: Using a time-series physiological data encoder to encode features of physiological state data can accurately capture the changing patterns of an individual's physiological state over time; a discrete behavior graph encoder to encode dynamic behavior feature data from campus cards and the library can deeply explore an individual's behavioral patterns on campus; a learning activity Transformer encoder to encode feature data from online learning platforms can comprehensively assess an individual's learning status; a static feature encoder to encode data from grade, major, and previous records from the psychological counseling center integrates the individual's basic feature information; by encoding different types of data separately, and then constructing a multimodal feature vector based on the generated physiological feature vector, behavioral feature vector, learning feature vector, and static feature vector, the fusion of multimodal information is achieved.

[0038] Example 6: A cross-modal attention fusion layer based on a dynamic hierarchical multi-task attention fusion model fuses and reduces the dimensionality of multimodal feature vectors to obtain a cross-modal fused feature vector, including: The behavioral feature vector, the learning feature vector, and the static feature vector are mapped to the same dimension as the physiological feature vector, and a feature matrix is ​​constructed with the physiological feature vector. Construct a grade-level guidance attention bias matrix; For each cross-attention head, generate a query matrix, a key matrix, and a value matrix; calculate the cross-attention score based on the query matrix and key matrix; normalize each cross-attention score using the Softmax function to obtain the attention weight corresponding to each cross-attention head; output the fusion feature of each cross-attention head based on the value matrix and attention weight of each cross-attention head. The fusion features of several cross-attention heads are concatenated into a multi-dimensional matrix, and the cross-modal fusion feature vector is obtained by dimensionality reduction through convolutional layers.

[0039] In this embodiment, a 1×1 convolutional layer maps the behavioral feature vector (256-dimensional), the learned feature vector (256-dimensional), and the static features (32-dimensional) to 512 dimensions, constructing a 4×512-dimensional feature matrix X with the physiological feature vector (512-dimensional); a grade-level guidance attention bias matrix B (4×4-dimensional) is also constructed: where This represents the grade-level adjustment weight of the i-th feature on the j-th feature; 8-head cross-attention calculation: For each attention head h (h=1~8), generate the query matrix Q_h=X×W_Qh, the key matrix K_h=X×W_Kh, and the value matrix V_h=X×W_Vh (W_Qh, W_Kh, and W_Vh are head-specific weight matrices, each with a dimension of 512×64); calculate the attention score S_h=(Q_h×K_h^T) / 8+B (adding the grade-level bias B, where T represents the matrix transpose); normalize S_h using the Softmax function to obtain the attention weight A_h; output the fusion feature V_h'=A_h×V_h for this head; feature concatenation and dimensionality reduction: concatenate the V_h' of the 8 heads (each with a dimension of 4×64) into a 4×512 dimensional matrix, and reduce the dimensionality to a 1×512 dimensional cross-modal fusion feature vector through a 1×1 convolutional layer.

[0040] The working principle and beneficial effects of the above technical solution are as follows: By mapping behavioral feature vectors, learned feature vectors, and static feature vectors to the same dimension as physiological feature vectors and constructing a feature matrix, feature information from different modalities can be fully integrated; the cross-modal attention fusion mechanism helps to discover potential relationships between features of different modalities; constructing a grade-guided attention bias matrix and calculating attention weights through a cross-attention mechanism enables the model to focus on important parts of different modal features according to different task requirements and data characteristics; the cross-attention mechanism can adaptively adjust the attention weights of different features according to the dynamic changes of input data. In practical applications, the distribution of data and the importance of features may change with time or scenario; by performing dimensionality reduction processing on the concatenated multidimensional matrix through convolutional layers to obtain cross-modal fusion feature vectors, the dimensionality of the data can be significantly reduced.

[0041] Example 7: Input the cross-modal fusion feature vector into the multi-task collaborative prediction layer of the dynamic hierarchical multi-task attention fusion model to predict the burnout risk level, and output the burnout risk level prediction result, including: The cross-modal fusion feature vector is input into the MLP network to obtain intermediate features. The intermediate features are divided into four sub-vectors: physiological factors, behavioral factors, learning factors, and grade factors. A risk factor attention pooling layer is constructed, and the attention weight of each sub-vector is calculated. The target intermediate vector is obtained by weighted summation based on the attention weights of each sub-vector. The target intermediate vector is input into the Softmax classification layer to obtain the burnout risk probability distribution. The risk level corresponding to the maximum burnout risk is used as the initial burnout prediction level, and the first loss function is calculated. Physiologically relevant sub-vectors are extracted from the cross-modal fusion feature vectors, input into the MLP network, outputting the sleep quality coefficient, and calculating the second loss function; Learn relevant sub-vectors are extracted from cross-modal fusion feature vectors, input into a 2-layer MLP network, output learning engagement coefficients, and a third loss function is calculated. Behavioral-physiological cross vectors are extracted from cross-modal fusion feature vectors, input into a 2-layer MLP network, output stress state coefficients, and a fourth loss function is calculated. Sleep quality coefficient, learning engagement coefficient, and stress state coefficient were used as auxiliary indicators. The target burnout risk level prediction results are determined based on the initial burnout prediction level, sleep quality coefficient, learning engagement coefficient, and stress state coefficient. A total loss function is constructed based on the first loss function, the second loss function, the third loss function, and the fourth loss function.

[0042] In this embodiment, the primary task is to predict the burnout risk level, while the secondary tasks are to predict sleep quality, learning engagement, and stress levels. The primary task (burnout risk level prediction) involves inputting a cross-modal fusion feature vector into a 3-layer MLP network (first layer dimension = 1024, second layer = 512, dropout probability = 0.3) to obtain a 512-dimensional intermediate feature F_main. F_main is then split into four 128-dimensional sub-vectors: physiological factor, behavioral factor, learning factor, and grade factor. The attention weights of each sub-vector (reflecting the factor's contribution to burnout) are calculated. A weighted sum is then obtained to obtain the target intermediate vector (512 dimensions), which is input into a Softmax classification layer (output dimension = 5, corresponding to burnout risk levels 0-4) to obtain the risk probability distribution P_burnout = [p0, p1, p2, p3, p4] and the predicted level R_burnout. The first loss function for the primary task is FocalLoss, with the formula: ,in The class weights are 0 = 0.5, 4 = 2.0, and γ = 2.0, representing the focusing parameter. This represents the predicted probability of category c; Auxiliary Task 1 (Sleep Quality Coefficient Prediction): Extract a 128-dimensional physiologically relevant sub-vector F_sleep from the cross-modal fusion feature vector, input it into a 2-layer MLP network (first layer dimension = 256, second layer = 1, activation function is Sigmoid), and output the sleep quality coefficient S_sleep (Sigmoid result × 10). The second loss function is... ; This represents the true sleep quality coefficient of the i-th sample; This represents the model's prediction coefficient for the sleep quality of the i-th sample; This indicates the number of samples involved in the calculation; Auxiliary Task 2 (Learning Engagement Coefficient Prediction): Extract a 128-dimensional learning-related sub-vector F_study from the cross-modal fusion feature vector, input it into a 2-layer MLP network (structure same as sleep quality coefficient prediction), and output the learning engagement coefficient S_study. The third loss function is... ; This represents the true learning engagement coefficient of the i-th sample; This represents the model's predicted learning engagement coefficient for the i-th sample; Auxiliary Task 3 (Stress State Coefficient Prediction): Extract a 128-dimensional behavior-physiology cross vector F_stress from the cross-modal fusion feature vector, input it into a 2-layer MLP network (structure as above), and output the stress state coefficient S_stress. The fourth loss function is... ;in, This represents the true pressure state coefficient of the i-th sample; The model's predicted stress state coefficients for the i-th sample represent the total loss function. in =1 indicates the loss weighting coefficient of the main task (burnout risk level prediction); , , These are the loss weighting coefficients for sleep quality, learning engagement, and stress state, respectively, and their sum is 1.

[0043] The working principle and beneficial effects of the above technical solution are as follows: By decomposing the cross-modal fusion feature vector into sub-vectors such as physiological factors, behavioral factors, learning factors, and grade factors, and calculating attention weights for each, the model can comprehensively consider information from multiple dimensions to predict the level of burnout risk; multi-task learning can share feature representations between different tasks, enabling the model to learn more general feature patterns from multiple related tasks; and constructing a total loss function based on the first, second, third, and fourth loss functions can simultaneously optimize the performance of multiple tasks. Different loss functions correspond to different tasks, and by combining them into a total loss function, the model can consider the optimization objectives of multiple tasks simultaneously during training, allowing each task to promote each other and jointly improve the overall performance of the model.

[0044] Example 8: Early warning module, including: The comparison submodule is used to compare the burnout risk level of the medical student to be evaluated with the preset warning threshold; The early warning submodule is used to issue an early warning when it is determined that the burnout risk level of the medical student to be evaluated is greater than the preset early warning threshold.

[0045] In this embodiment, the preset warning threshold is determined based on experimental data.

[0046] Example 9: Intervention module, including: The query submodule is used to query the risk level-intervention strategy table based on the burnout risk level of the medical student to be evaluated, and to determine the target intervention strategy. The intervention submodule is used to intervene in medical students under assessment based on goal-oriented intervention strategies.

[0047] In this embodiment, specific intervention measures in the intervention strategy include: 1. Pushing mindfulness breathing guidance videos to students and recommending psychological counseling services; 2. Sending anonymous risk warnings to counselors or mentors and suggesting care talks; 3. Intelligently adjusting the pace and difficulty of task pushes to the student in the learning system.

[0048] Example 10: It also includes an intervention assessment module, which is used to obtain the change index of the burnout risk level of the medical students within the preset time period after the intervention strategy is implemented on the medical students, and to dynamically adjust the intervention strategy for the medical students based on the change index.

[0049] The working principle and beneficial effects of the above technical solution are as follows: By comparing the burnout risk level of the medical students to be assessed with the preset warning threshold through the comparison submodule, it is possible to detect situations where the burnout risk of medical students exceeds the normal range in the first instance; medical students face considerable pressure during their studies and practice, and burnout may quietly develop. Early detection allows students to understand their own state in a timely manner, preventing the burnout from worsening and protecting their physical and mental health; the query submodule determines the target intervention strategy based on the burnout risk level query risk level-intervention strategy table, and the intervention submodule implements the intervention accordingly; different burnout risk levels correspond to different intervention strategies, providing personalized solutions for medical students. For example, for students with mild burnout, gentler intervention methods such as psychological counseling lectures and group discussions may be used; while for students with severe burnout, more in-depth intervention measures such as one-on-one psychological counseling and adjusting study plans may be provided to improve the effectiveness of the intervention; after the preset time period of implementing the intervention strategy, the change index of the burnout risk level is obtained, and the intervention strategy is dynamically adjusted; the burnout situation of medical students is dynamic, and different stages may require different intervention methods. Dynamic adjustments can ensure that intervention strategies always adapt to students' actual situations, continuously and effectively help students alleviate burnout, and promote their better engagement in learning and practice.

[0050] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A medical student burnout risk early warning and intervention system based on dynamic data monitoring, characterized in that, include: The acquisition module is used to acquire dynamic monitoring data and static basic data of the medical students to be evaluated. The preprocessing module is used to preprocess the dynamic monitoring data to obtain preprocessed dynamic monitoring data; The prediction module is used to input static basic data and preprocessed dynamic monitoring data into a pre-trained burnout risk prediction model for prediction, and output the burnout risk level of the medical student to be evaluated. The early warning module is used to determine whether to issue an early warning based on the burnout risk level of the medical student to be assessed. The intervention module is used to obtain the intervention strategy corresponding to the burnout risk level of the medical student to be evaluated when it is determined to issue an early warning prompt, and to intervene on the medical student to be evaluated based on the intervention strategy.

2. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 1, characterized in that, The acquisition module includes: The first acquisition submodule is used to acquire dynamic behavioral data of the medical students to be evaluated; the dynamic behavioral data includes campus card usage record data, library behavior data, and online learning platform data. The second acquisition submodule is used to acquire physiological status data of the medical student to be evaluated; the physiological status data includes sleep structure data, activity level data, and heart rate variability data. The determination submodule is used to use the dynamic behavioral data and physiological status data of the medical students to be evaluated as dynamic monitoring data of the medical students to be evaluated. The third acquisition submodule is used to acquire the static basic data of the medical students to be evaluated; the static basic data includes the grade, major, and previous psychological counseling center records of the medical students to be evaluated.

3. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 1, characterized in that, The preprocessing module includes: The data cleaning submodule is used to clean the dynamic monitoring data to obtain cleaned dynamic monitoring data. The noise reduction submodule is used to reduce the noise of the cleaned dynamic monitoring data to obtain processed dynamic monitoring data.

4. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 1, characterized in that, Training methods for burnout risk prediction models include: Obtain the training dataset for the burnout risk prediction model; Feature extraction is performed on the training dataset to obtain the multimodal training features corresponding to the training dataset; The multimodal training features are input into the multi-source data feature encoder of the dynamic hierarchical multi-task attention fusion model for feature encoding, and the multimodal training features are converted into multimodal feature vectors. The cross-modal attention fusion layer based on the dynamic hierarchical multi-task attention fusion model fuses and reduces the dimensionality of multimodal feature vectors to obtain cross-modal fused feature vectors. The cross-modal fusion feature vector is input into the multi-task collaborative prediction layer of the dynamic hierarchical multi-task attention fusion model to predict the burnout risk level, and the burnout risk level prediction result is output. The training dataset of the burnout risk prediction model is sequentially input into the dynamic hierarchical multi-task attention fusion model for training. When the training results meet the requirements, the initial burnout risk prediction model is obtained. Obtain the test dataset for the burnout risk prediction model; The initial burnout risk prediction model was tested using a test dataset based on the burnout risk prediction model. When the test results met the requirements, the trained burnout risk prediction model was obtained.

5. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 4, characterized in that, The multimodal training features are input into the multi-source data feature encoder of the dynamic hierarchical multi-task attention fusion model for feature encoding, converting the multimodal training features into multimodal feature vectors, including: Based on the time-series physiological data encoder, the physiological state data included in the multimodal training features are feature-encoded to generate physiological feature vectors; Based on the discrete behavior graph encoder, feature encoding is performed on the campus card and library dynamic behavior feature data included in the multimodal training features to generate behavior feature vectors; Based on the learning activity Transformer encoder, feature encoding is performed on the online learning platform feature data included in the multimodal training features to generate learning feature vectors; Based on the static feature encoder, the grade, major and previous psychological counseling center records included in the multimodal training features are feature encoded to generate static feature vectors; Multimodal feature vectors are constructed based on physiological feature vectors, behavioral feature vectors, learning feature vectors, and static feature vectors.

6. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 5, characterized in that, A cross-modal attention fusion layer based on a dynamic hierarchical multi-task attention fusion model fuses and reduces the dimensionality of multimodal feature vectors to obtain cross-modal fused feature vectors, including: The behavioral feature vector, the learning feature vector, and the static feature vector are mapped to the same dimension as the physiological feature vector, and a feature matrix is ​​constructed with the physiological feature vector. Construct a grade-level guidance attention bias matrix; For each cross-attention head, generate a query matrix, a key matrix, and a value matrix; calculate the cross-attention score based on the query matrix and key matrix; normalize each cross-attention score using the Softmax function to obtain the attention weight corresponding to each cross-attention head; output the fusion feature of each cross-attention head based on the value matrix and attention weight of each cross-attention head. The fusion features of several cross-attention heads are concatenated into a multi-dimensional matrix, and the cross-modal fusion feature vector is obtained by dimensionality reduction through convolutional layers.

7. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 6, characterized in that, The cross-modal fusion feature vector is input into the multi-task collaborative prediction layer of the dynamic hierarchical multi-task attention fusion model to predict the burnout risk level, and the burnout risk level prediction results are output, including: The cross-modal fusion feature vector is input into the MLP network to obtain intermediate features. The intermediate features are divided into four sub-vectors: physiological factors, behavioral factors, learning factors, and grade factors. A risk factor attention pooling layer is constructed, and the attention weight of each sub-vector is calculated. The target intermediate vector is obtained by weighted summation based on the attention weights of each sub-vector. The target intermediate vector is input into the Softmax classification layer to obtain the burnout risk probability distribution. The risk level corresponding to the maximum burnout risk is used as the initial burnout prediction level, and the first loss function is calculated. Physiologically relevant sub-vectors are extracted from the cross-modal fusion feature vectors, input into the MLP network, outputting the sleep quality coefficient, and calculating the second loss function; Learn relevant sub-vectors are extracted from cross-modal fusion feature vectors, input into a 2-layer MLP network, output learning engagement coefficients, and a third loss function is calculated. Behavioral-physiological cross vectors are extracted from cross-modal fusion feature vectors, input into a 2-layer MLP network, output stress state coefficients, and a fourth loss function is calculated. Sleep quality coefficient, learning engagement coefficient, and stress state coefficient were used as auxiliary indicators. The target burnout risk level prediction results are determined based on the initial burnout prediction level, sleep quality coefficient, learning engagement coefficient, and stress state coefficient. A total loss function is constructed based on the first loss function, the second loss function, the third loss function, and the fourth loss function.

8. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 1, characterized in that, The early warning module includes: The comparison submodule is used to compare the burnout risk level of the medical student to be evaluated with the preset warning threshold; The early warning submodule is used to issue an early warning when it is determined that the burnout risk level of the medical student to be evaluated is greater than the preset early warning threshold.

9. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 1, characterized in that, Intervention modules, including: The query submodule is used to query the risk level-intervention strategy table based on the burnout risk level of the medical student to be evaluated, and to determine the target intervention strategy. The intervention submodule is used to intervene in medical students under assessment based on goal-oriented intervention strategies.

10. The medical student burnout risk early warning and intervention system based on dynamic data monitoring as described in claim 1, characterized in that, It also includes an intervention assessment module, which is used to obtain the change index of the burnout risk level of the medical students within the preset time period after the intervention strategy is implemented on the medical students, and to dynamically adjust the intervention strategy for the medical students based on the change index.

Images