A method and system for dynamically constructing a nursing student learning portrait based on multi-source data fusion

Abstract

The application discloses a kind of based on multi-source data fusion's nursing student learning portrait dynamic construction method and system, relating to nursing education technical field, the method includes: to the feature screening of nursing student learning portrait, selects the core anchoring feature dimension and auxiliary check dimension of nursing student learning portrait;Quality check and deviation calibration are carried out to the collected data, and output standardization effective feature set;Fusion feature vector is input into empirical calibration classification model and empirical weight calibration prediction model, generates and stores initial nursing student learning portrait;The dynamic updating mechanism of time and event driving is constructed;Based on the potential category attribution of dynamically updated nursing student learning portrait, the corresponding intervention push and risk early warning are matched.The application realizes the accurate locking of nursing student learning portrait core dimension and the empirical calibration of corresponding weight, breaks through the application limitation of conventional experience type feature selection, makes the strong association of portrait core dimension and nursing student learning result.

Description

Technical Field

[0001] This invention relates to the field of nursing education technology, and in particular to a method and system for dynamically constructing learning profiles of nursing students based on multi-source data fusion. Background Technology

[0002] Nursing, a profession that combines theoretical and practical aspects with professional ethics, has seen blended learning, integrating online and offline methods, become the mainstream approach in nursing education to meet talent development needs. Online self-study, a core component of blended learning, directly determines students' theoretical mastery, self-learning abilities, and ultimately, the overall effectiveness of nursing talent development. Learning profiling technology, through systematic analysis of learning process data, can reveal learners' learning status, abilities, and development potential. It serves as a core technological vehicle for personalized learning guidance and precise teaching intervention for nursing students. Optimization of profiling technology for nursing education scenarios possesses clear industry application value and practical significance for teaching.

[0003] After years of development, learning profiling technology has formed a relatively mature technological evolution path. Early technology focused on statistical analysis of user behavior on online learning platforms, primarily visualizing basic indicators such as learning duration and task completion rates. With the popularization of machine learning technology, profiling technology has gradually evolved into a comprehensive modeling solution that integrates multi-source data. Through algorithms such as clustering and regression, it achieves functions such as learner classification, learning outcome prediction, and risk warning. It has been implemented on a large scale in general education online education and higher education scenarios, and its technological system is becoming increasingly sophisticated.

[0004] However, in the specific context of nursing education, the application of existing general learning profiling technologies still faces several limitations. Most existing solutions follow the general technical framework of general education, with feature selection primarily based on data availability and online learning behavior data as the core modeling dimension. This results in insufficient exploration of the learning characteristics of nursing students and the core factors influencing long-term learning outcomes, leading to a weak correlation between the profiling results and students' actual learning performance. At the model construction level, existing technologies often use default algorithm parameters or conventional cross-validation optimization, lacking systematic empirical calibration with large-sample data from nursing students. The model outputs often only possess mathematical statistical significance, lacking interpretability adapted to nursing teaching scenarios, and exhibiting unstable generalization performance across different institutions and academic programs.

[0005] Meanwhile, existing technologies often use a fixed-cycle full-update mode for dynamic updates, which can easily lead to meaningless fluctuations in profiles and redundancy in system resources. For subjective data such as self-assessment by nursing students, quality control is only achieved by eliminating invalid basic responses, lacking a systematic bias identification and correction mechanism, which can easily affect the final accuracy of the profile due to data bias. Summary of the Invention

[0006] Therefore, it is necessary to provide a method and system for dynamically constructing nursing student learning profiles based on multi-source data fusion to address the aforementioned technical problems.

[0007] In a first aspect, the present invention provides a method for dynamically constructing nursing student learning profiles based on multi-source data fusion, including:

[0008] S1. Based on the full-dimensional feature dataset of nursing students, feature selection is performed on the nursing student learning profile. The core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile are selected, and the fixed weights of each dimension are empirically calibrated to construct an empirical calibration classification model and an empirical weight calibration prediction model.

[0009] S2. Perform two-dimensional hierarchical data collection based on the core anchoring feature dimension and the auxiliary verification dimension, and perform quality verification and deviation calibration on the collected data to output a standardized effective feature set.

[0010] S3. Using the fixed weights that have been labeled, perform weighted deep fusion on the standardized effective feature set to generate a fusion feature vector corresponding to a single nursing student. Then, input the fusion feature vector into the empirical labeling classification model and the empirical weight labeling prediction model to generate and store the initial nursing student learning profile.

[0011] S4. Based on the core anchor feature dimension, a time- and event-driven dynamic update mechanism is constructed. When the update threshold is reached, the nursing student learning profile is updated, and the profile change trajectory is retained.

[0012] S5. Based on the potential category attribution of the dynamically updated nursing student learning profile, match the corresponding intervention push and risk warning, collect intervention effect feedback data, and perform reverse verification on the fixed weights.

[0013] Furthermore, based on the full-dimensional feature dataset of nursing students, feature selection was performed on the nursing student learning profiles. Core anchoring feature dimensions and auxiliary verification dimensions were selected, and fixed weights for each dimension were empirically calibrated to construct an empirically calibrated classification model and an empirically calibrated prediction model, including:

[0014] S11. Using a fully relevant feature selection algorithm combined with interpretability analysis, effective features in the full-dimensional feature dataset of nursing students are selected, and the marginal contribution of each effective feature to the prediction of nursing students' learning results is quantified. Based on the marginal contribution results, the core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile are divided, and fixed weights are set for each dimension.

[0015] S12. Based on the core anchoring feature dimension, auxiliary verification dimension and fixed weight, the potential profile analysis pre-training is carried out using the nursing student effective labeled dataset, and the accuracy is verified by the classification threshold. After the verification is passed, it is solidified into the empirical labeling classification model of potential profile analysis.

[0016] S13. Based on the core anchoring feature dimension, auxiliary verification dimension and fixed weights, the nursing student assessment score is used as the output label. Generalized linear regression is pre-trained using an effective labeled dataset. Empirical calibration is performed by combining marginal contribution degree to obtain the empirical weight calibration prediction model.

[0017] Furthermore, a fully relevant feature selection algorithm combined with interpretability analysis was used to select effective features from the full-dimensional feature dataset of nursing students, and the marginal contribution of each effective feature to the prediction of nursing student learning outcomes was quantified. Based on the marginal contribution results, the core anchoring feature dimension and auxiliary verification dimension of the nursing student learning profile were divided, and fixed weights were set for each dimension, including:

[0018] S111. Create randomly permuted shaded features for the original features in the nursing student full-dimensional feature dataset, and perform two-sided statistical verification on the feature importance of the original features and the shaded features; based on the preset significance threshold, iteratively filter all effective features that are statistically correlated with the nursing student learning results.

[0019] S112. The marginal contribution of all effective features to the prediction of nursing students' learning outcomes is calculated using an interpretability algorithm. The features are sorted from high to low according to their marginal contribution. The online learning patience scale and the exploration community scale, which have the highest comprehensive interpretability contribution, are selected as the core anchor feature dimensions for the nursing students' learning profile. The remaining effective features are divided into auxiliary verification dimensions according to their contribution attributes.

[0020] S113. Based on the interpretability contribution ratio of each feature, empirically calibrate the fixed weights of each feature within the core anchor feature dimension and the auxiliary verification dimension.

[0021] S114. Digitize and standardize the core anchoring feature dimensions, select the hierarchical scoring rules, total score range, and reverse scoring conversion rules for the online learning patience scale, and select the hierarchical scoring rules, total score range, and positive scoring conversion rules for the exploration community scale.

[0022] Furthermore, a pre-training process for latent profile analysis was conducted using a nursing student-labeled dataset, and accuracy was verified using a classification threshold. Once verified, the model was solidified into an empirically calibrated classification model for latent profile analysis, including:

[0023] S121. Select several sets of effective labeled datasets of nursing students collected in actual practice as pre-training samples. Take the standardized feature data of the core anchoring feature dimension and auxiliary verification dimension in the pre-training samples as input and substitute them into the potential profile analysis model for pre-training.

[0024] S122. The model fitting effect of different numbers of categories is compared one by one through the likelihood ratio test to select the optimal number of categories. The characteristic discrimination of each category of nursing students is analyzed by the receiver operating characteristic curve (ROC curve). Combined with the Youden index calculation results, the optimal classification threshold range of each category of nursing students is selected.

[0025] S123. Preset the fitting hard threshold of the empirical calibration classification model for latent profile analysis, and calculate the classification accuracy of the model through the confusion matrix. Perform multiple rounds of accuracy verification on the pre-trained latent profile analysis model. When the accuracy meets the preset threshold, the solidified empirical calibration classification model is obtained.

[0026] Furthermore, using nursing students' assessment scores as output labels, a generalized linear regression pre-training process is performed using an effectively labeled dataset. Empirical calibration is then performed using marginal contribution, resulting in an empirical weighting prediction model that includes:

[0027] S131. Using the theoretical and practical assessment scores of nursing students as the output labels, select the effective labeled dataset as the pre-training samples, take the full standardized feature data in the pre-training samples as the input, and use the multiple linear regression algorithm to build the initial prediction model and perform pre-training.

[0028] S132. Combine the marginal contribution of each feature with the use of interpretable algorithms to perform empirical calibration of the weight coefficients of the initial prediction model in order to calibrate the fixed weight coefficients of each feature.

[0029] S133. Based on the lower limit of the confidence interval of the pre-training results, set a hard threshold for the goodness of fit of the generalized linear empirical weight calibration prediction model, and calculate the average absolute error between the model prediction value and the actual assessment score of nursing students. Verify the accuracy of the pre-trained model. When the goodness of fit and the average absolute error both meet the preset requirements, the solidified empirical weight calibration prediction model is obtained.

[0030] Furthermore, the standardized effective feature set is weighted and deeply fused using the pre-defined fixed weights to generate a fused feature vector corresponding to each nursing student. This fused feature vector is then input into the empirically calibrated classification model and the empirically weighted prediction model to generate and store the initial nursing student learning profile, including:

[0031] S31. Perform weighted feature fusion on the standardized effective feature set according to the fixed weights, and generate a unique one-dimensional fusion feature vector corresponding to each nursing student.

[0032] S32. Calculate the Pearson correlation coefficient between a single fused feature vector and the nursing student's historical learning results, and compare it with the preset correlation coefficient threshold. Automatically return samples that fail the verification to re-execute data collection, quality verification and deviation calibration.

[0033] S33. Input the validated fusion feature vector into the empirical calibration classification model to output the potential category affiliation and classification confidence of a single nursing student; and input the fusion feature vector into the empirical weight calibration prediction model to output the predicted value of the nursing student's learning result, and calibrate the learning risk level based on the predicted value.

[0034] S34. Map the potential category affiliation, predicted result value and standardized feature data of nursing students into multi-layer structured labels, and set a quantitative value range for each label; generate an initial nursing student learning profile corresponding to a single nursing student based on the full-dimensional labels, and store it in the profile database after associating it with the nursing student's anonymous ID.

[0035] Furthermore, based on the core anchored feature dimension, a time- and event-driven dynamic update mechanism is constructed. When the update threshold is reached, the nursing student learning profile is updated, and the profile change trajectory is retained, including:

[0036] S41. Based on the longitudinal tracking data of the actual collected nursing student samples, the natural fluctuation range of nursing students' online learning patience is statistically obtained, and the change of the online learning patience scale score is greater than or equal to the preset threshold as the substantive change threshold of the core anchor feature dimension.

[0037] S42. Set time-driven anchor point review updates to match the blended teaching cycle of nursing students. Preset multiple fixed anchor point review nodes. When a fixed node is reached, push the online learning patience scale and the inquiry community scale. When the change in the retest score is greater than or equal to the substantive change threshold, trigger the full collection of auxiliary verification dimension data and the full-dimensional update of nursing student learning profile.

[0038] S43. Set up event-driven anchor point linkage update, preset multiple types of pre-trigger conditions. When any pre-trigger event is triggered, first perform the core anchor dimension retest. When the retest score is greater than or equal to the substantial change threshold, trigger the full-dimensional data collection and nursing student learning profile update.

[0039] S44. When the number of newly added valid labeled samples in the portrait database reaches the initial pre-training sample size, the model retraining process is triggered, and the newly added valid samples are added to the training set to re-optimize the model parameters of the empirical labeling classification model and the empirical weight labeling prediction model.

[0040] S45. Synchronously store the dynamically updated portrait data and model iteration records in the portrait database, and retain the portrait change trajectory throughout the entire learning cycle of nursing students.

[0041] Secondly, a dynamic construction system for nursing student learning profiles based on multi-source data fusion is provided, which includes:

[0042] The baseline model building module is used to perform feature selection on the nursing student learning profile based on the nursing student full-dimensional feature dataset, select the core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile, and perform empirical calibration on the fixed weight of each dimension to build an empirical calibration classification model and an empirical weight calibration prediction model.

[0043] The data acquisition and recognition module is used to perform two-dimensional hierarchical data acquisition based on the core anchor feature dimension and the auxiliary verification dimension, and to perform quality verification and deviation calibration on the acquired data, and output a standardized and effective feature set.

[0044] The profile fusion generation module is used to perform weighted deep fusion on the standardized effective feature set using the labeled fixed weights, generate a fusion feature vector corresponding to a single nursing student, and input the fusion feature vector into the empirical labeling classification model and the empirical weight labeling prediction model to generate and store the initial nursing student learning profile.

[0045] The dynamic iterative update module is used to build a time- and event-driven dynamic update mechanism based on the core anchor feature dimension. When the update threshold is reached, the nursing student learning profile is updated and the profile change trajectory is preserved.

[0046] The intervention closed-loop verification module is used to match the corresponding intervention push and risk warning based on the potential category classification of the dynamically updated nursing student learning profile, and to collect intervention effect feedback data to perform reverse verification on fixed weights.

[0047] Thirdly, the present invention provides an electronic device, which includes a processor, a storage medium and a computer program, wherein the computer program is stored in the storage medium, and when the computer program is executed by the processor, it implements the above-mentioned method for dynamically constructing nursing student learning profiles based on multi-source data fusion.

[0048] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for dynamically constructing nursing student learning profiles based on multi-source data fusion.

[0049] The beneficial effects of this invention are as follows:

[0050] 1. By combining fully relevant feature screening with interpretability analysis, the core dimensions of nursing student learning profiles were accurately identified and their corresponding weights were empirically calibrated. This approach breaks through the limitations of conventional empirical feature selection and establishes a strong correlation between the core dimensions of the profiles and the nursing student learning outcomes. At the same time, the fixed weight design ensures the consistency of the profile benchmark and avoids the problem of core features being submerged by noise due to automatic weight optimization by the algorithm, effectively improving the group adaptability and result stability of the profiles.

[0051] 2. By constructing a classification and prediction model adapted to the characteristics of nursing students, and combining it with a dual-mode dynamic update mechanism triggered by core anchor points, the core problems of insufficient model generalization and redundant update operations in conventional technologies are solved. This ensures that the profile can accurately capture the substantial changes in the learning status of nursing students, while avoiding the system resource consumption caused by meaningless full updates, thus achieving a two-way balance between profile recognition accuracy and system operating efficiency.

[0052] 3. By identifying and calibrating the contradictions between self-assessment data and behavioral data, the whole-process quality control system for subjective data has been improved, effectively enhancing the effectiveness of the profile input data. At the same time, by combining the empirical intervention matching of profile classification results and the reverse optimization design of intervention effect feedback, a complete technical closed loop from profile construction to teaching implementation has been formed, making the profile technology truly adaptable to the personalized training needs of nursing blended teaching and greatly enhancing the practical application value of the technology. Attached Figure Description

[0053] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0054] Figure 1 This is a flowchart of a method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to an embodiment of the present invention;

[0055] Figure 2 This is a flowchart of step S3, generating a nursing student learning profile, in a method for dynamically constructing a nursing student learning profile based on multi-source data fusion according to an embodiment of the present invention.

[0056] Figure 3 This is a line graph showing the score distribution of 10 items of the SOLPS scale for three potential categories of nursing students in a dynamic construction method for nursing student learning profiles based on multi-source data fusion according to an embodiment of the present invention.

[0057] Figure 4 This is a principle block diagram of a dynamic construction system for nursing student learning profiles based on multi-source data fusion according to an embodiment of the present invention;

[0058] Figure 5 This is a schematic block diagram of an electronic device according to an embodiment of the present invention.

[0059] The diagram shows the following modules: 1. Baseline model construction module; 2. Data acquisition and recognition module; 3. Profile fusion and generation module; 4. Dynamic iterative update module; 5. Intervention closed-loop verification module. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] Please see Figure 1 This paper presents a method for dynamically constructing nursing student learning profiles based on multi-source data fusion, including:

[0062] S1. Based on the full-dimensional feature dataset of nursing students, feature selection is performed on the nursing student learning profile. The core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile are selected, and the fixed weights of each dimension are empirically calibrated to construct an empirical calibration classification model and an empirical weight calibration prediction model.

[0063] In the description of this invention, based on a full-dimensional feature dataset of nursing students, feature selection is performed on the nursing student learning profile, selecting the core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile, and empirically calibrating the fixed weights of each dimension to construct an empirically calibrated classification model and an empirically calibrated prediction model, including:

[0064] S11. Using a fully relevant feature selection algorithm combined with interpretability analysis, effective features in the full-dimensional feature dataset of nursing students are selected, and the marginal contribution of each effective feature to the prediction of nursing students' learning results is quantified. Based on the marginal contribution results, the core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile are divided, and fixed weights are set for each dimension.

[0065] In this invention, a fully relevant feature selection algorithm combined with interpretability analysis is used to select effective features from a full-dimensional feature dataset of nursing students, and the marginal contribution of each effective feature to the prediction of nursing student learning outcomes is quantified. Based on the marginal contribution results, the core anchoring feature dimension and auxiliary verification dimension of the nursing student learning profile are divided, and fixed weights are set for each dimension, including:

[0066] S111. Create randomly permuted shadow features for the original features in the nursing student full-dimensional feature dataset, and perform two-sided statistical verification on the feature importance of the original features and shadow features; based on the preset significance threshold, iteratively screen all effective features that are statistically correlated with the nursing student learning results.

[0067] Specifically, this invention employs the Boruta fully relevant feature selection algorithm to address the technical problems of existing recursive feature elimination (RFE) algorithms, such as high false positive rates, inability to capture inter-feature interaction effects, and susceptibility to false rejection of core features. The base classifier of the fully relevant feature selection algorithm uses a random forest model with a structure of 500 decision trees, a maximum tree depth of 10 layers, a minimum number of samples per leaf node of 5, and a minimum impurity reduction threshold of 1e-7 for node splitting. Out-of-bag (OOB) data is used to verify feature importance.

[0068] The shadow feature generation rule involves randomly shuffling the values ​​of each original feature without replacement to generate shadow features that have the same distribution as the original features but are unrelated to the students' learning results. Each shadow feature corresponds one-to-one with an original feature and participates in model training simultaneously. In the feature importance calculation logic, the feature importance of each true feature and shadow feature is calculated based on the Gini coefficient. The maximum importance of all shadow features is used as a threshold; true features with importance higher than this threshold are marked as "important features," while those lower are marked as "unimportant features."

[0069] This invention employs a Z-test to perform two-sided statistical verification of the importance mean of true and shading features, with a preset significance threshold of P<0.05 to strictly control Type I errors. The number of iterations is fixed at 100, continuously eliminating statistically insignificant redundant features during iteration, and retaining all valid features after iteration termination. The number of decision trees is 500; the 100 iterations keep the false positive rate below 2%, adjustable from 50 to 200 iterations; the maximum tree depth is 10 layers.

[0070] S112. The marginal contribution of all effective features to the prediction of nursing students' learning outcomes is calculated using an interpretability algorithm. The features are sorted from high to low according to their marginal contribution. The online learning patience scale and the exploration community scale, which have the highest comprehensive interpretability contribution, are selected as the core anchor feature dimensions for the nursing student learning profile. The remaining effective features are divided into auxiliary verification dimensions according to their contribution attributes.

[0071] Specifically, this step employs an interpretability algorithm (Tree-SHAP algorithm: SHapley Additive ex Planations, a tree-based Shapley additive interpretation method) to address the technical problems of existing feature importance assessment methods lacking consistency and failing to accurately quantify the independent marginal contribution of individual features. The base model uses an XGBoost regression tree model with a structure of 300 regression trees, a maximum tree depth of 6 layers, a learning rate of 0.01, an L2 regularization coefficient λ=0.01, and a quadratic squared loss function. The final comprehensive assessment score for nursing students is used as the output label.

[0072] Based on the Shapley value principle in cooperative game theory, the average marginal contribution of each feature to the prediction result under all feature combinations is calculated using the following formula:

[0073] ;

[0074] in, SHAP value (marginal contribution) of the i-th feature. For all feature sets, For feature subset, For the prediction model; ! represents the standard factorial symbol; Input all features for a single sample.

[0075] The overall interpretability contribution of a feature is calculated by summing the absolute values ​​of its SHAP values ​​to the total SHAP values. After sorting the contributions from highest to lowest, the combined contribution of the 10 items of the Online Learning Patience Scale (SOLPS) and the 32 items of the Inquiry Community Scale (CII) reached 75%, and were therefore selected as the core anchoring feature dimensions. The remaining effective features were divided into three categories of auxiliary verification dimensions based on their attributes: basic attribute dimensions, online learning behavior dimensions, and learning outcome dimensions. The basic attribute dimensions include gender, school system, and grade level. The online learning behavior dimensions include video playback interruption rate, task completion time, and interaction participation. The learning outcome dimensions include historical assessment scores.

[0076] The interpretability algorithm uses 300 regression trees to ensure the stability of SHAP value calculation, with an adjustment range of 100-500 trees; a learning rate of 0.01 to balance convergence speed and fitting accuracy, with an adjustment range of 1e-30.1; and an L2 regularization coefficient of 0.01 to control the risk of overfitting, with an adjustment range of 1e-41e-1.

[0077] S113. Based on the interpretability contribution ratio of each feature, empirically calibrate the fixed weights of each feature within the core anchor feature dimension and the auxiliary verification dimension.

[0078] Specifically, the fixed weighting rule is as follows: the fixed weight of a single feature = the absolute value of the feature's SHAP marginal contribution / the sum of the absolute values ​​of the SHAP marginal contributions of all valid features, with the total weight summed to 1. The weights of each item within the core anchor feature dimension are allocated within 75% of the total weight based on its own SHAP contribution, with the SOLPS core feature weighting at 42% and the CII core feature weighting at 33%. The weights of each feature within the auxiliary validation dimension are allocated within 25% of the total weight based on its own SHAP contribution. Once the weights are calibrated, they remain fixed throughout the process and are not automatically adjusted with single-sample data.

[0079] Weight adjustment rules: Weight recalibration is triggered only when the absolute value of the SHAP contribution of a certain feature changes by ≥0.2 in the subsequent full validation data; otherwise, it remains fixed.

[0080] This step avoids the problem of core features being overwhelmed by noise due to the automatic learning of weights in conventional algorithms, ensuring that the profile is always anchored to the core psychological traits of protecting life.

[0081] S114. Digitize and standardize the core anchoring feature dimensions, select the hierarchical scoring rules, total score range, and reverse scoring conversion rules for the online learning patience scale, and select the hierarchical scoring rules, total score range, and positive scoring conversion rules for the exploration community scale.

[0082] Specifically, the Online Learning Patience Scale (SOLPS) is a 10-item standardized scale that has been localized and tested for reliability and validity. The scoring rule adopts a Likert 5-point scale, where 1 point = strongly disagree, 2 points = disagree, 3 points = uncertain, 4 points = agree, and 5 points = strongly agree. The total score ranges from 10 to 50 points. The reverse scoring conversion rule is as follows: the 6 reverse-scored items in the scale are converted by "6 ​​- original score". The lower the total score after conversion, the higher the nursing student's level of online learning patience.

[0083] The Community Inquiry Scale (CII) is a standardized 32-item scale with 4 dimensions, validated for reliability and validity among nursing students. The 4 dimensions are cognitive presence, teaching presence, social presence, and learning presence. The scoring rule adopts a Likert 5-point scale, where 1 point = completely disagree, 2 points = mostly disagree, 3 points = uncertain, 4 points = mostly agree, and 5 points = completely agree. The total score ranges from 32 to 160 points. The positive scoring conversion rule is to directly add up the scores of all items. The higher the total score, the better the nursing students' blended learning experience.

[0084] This step completes the digital encapsulation of the psychological scale, forming a standardized feature field that can be automatically processed by computers. This solves the problem that scale data cannot be directly used for model calculation and ensures that the feature data of different samples are comparable across different samples.

[0085] S12. Based on the core anchoring feature dimension, auxiliary verification dimension and fixed weight, the potential profile analysis pre-training is carried out using the nursing student effective labeled dataset, and the accuracy is verified by the classification threshold. After the verification is passed, it is solidified into the empirical labeling classification model of potential profile analysis.

[0086] In the description of this invention, a latent profile analysis pre-training is performed using a nursing student effectively labeled dataset, and the accuracy is verified by a classification threshold. After passing the verification, the model is solidified into an empirically labeled classification model for latent profile analysis, including:

[0087] S121. Select several sets of effective labeled datasets of nursing students collected in practice as pre-training samples. Take the standardized feature data of the core anchoring feature dimension and auxiliary verification dimension in the pre-training samples as input and substitute them into the potential profile analysis model for pre-training.

[0088] Specifically, this step uses the Latent Profile Analysis (LPA) model to address the technical problems of existing K-means hard clustering algorithms, such as the inability to output probabilistic interpretations, poor adaptability to non-normally distributed data, and the lack of educational and psychological meaning in the classification results.

[0089] The pre-training sample consisted of valid labeled data from 764 nursing students in three medical colleges in Heilongjiang, Liaoning, and Shandong provinces. The sample covered 412 four-year undergraduate students and 352 three-year vocational college students, spanning their first to third year of study, with a male-to-female ratio of 1:8.2, consistent with the student demographics of nursing programs in China. All students signed electronic informed consent forms. Data underwent complete anonymization and desensitization, with unique identifiers such as names and student IDs replaced by irreversible hash values, and sensitive information such as ID numbers and contact information completely removed. The dataset was partitioned using stratified sampling, stratified by academic system and year of study, dividing the 764 samples into a 7:1:2 ratio: 535 samples for training, 76 samples for validation, and 153 samples for testing. A fixed random seed of 20240601 was used to ensure experimental reproducibility.

[0090] Based on the theory of finite mixture models, it is assumed that the features within each potential class follow a multivariate normal distribution. The robust maximum likelihood (MLR) method is used for parameter estimation to avoid fitting bias caused by non-normal data distribution. The model input consists of standardized core anchor features and auxiliary verification features, and the output is the posterior probability of each potential class for each sample.

[0091] The standardized feature data of the training set is input into the model, and the potential profile models of classes 1-5 are fitted sequentially. The EM algorithm is used to solve the parameters iteratively. The maximum number of iterations is preset to 10,000, and the convergence threshold is 1e-8. The iteration stops when the change in the log-likelihood value between two iterations is less than the convergence threshold.

[0092] S122. The model fitting effect of different numbers of categories is compared one by one through the likelihood ratio test to select the optimal number of categories. The characteristic discrimination of each category of nursing students is analyzed by the receiver operating characteristic curve (ROC curve). Combined with the Youden index calculation results, the optimal classification threshold range of each category of nursing students is selected.

[0093] Specifically, after fitting potential profile models of classes 1-5 sequentially, the goodness of fit of the adjacent number of classes models was compared using the Lo-Mendell-Rubin likelihood ratio test (LMR) and the Bootstrap likelihood ratio test (BLRT). When P < 0.001, it means that increasing the number of classes can significantly improve the model fitting effect. The final validation results show that when the number of classes is 3, the P values ​​of both LMR and BLRT tests are < 0.001, and when the number of classes is increased to 4, the P value is > 0.05. Therefore, the optimal number of classes is determined to be 3.

[0094] For the three potential categories, ROC curves were plotted using the total SOLPS score and the total CII score as test variables, respectively. Sensitivity and specificity were calculated at different thresholds. The Youden index was calculated using the formula: Youden index = sensitivity + specificity - 1. The threshold corresponding to the maximum value of the Youden index is the optimal classification threshold.

[0095] The three threshold ranges were finally determined as follows: High patience-focused type, SOLPS total score ≤ 22 points, CII total score ≥ 112 points, Youden index 0.94; Medium patience-fluctuation type, SOLPS total score 23-38 points, CII total score 89-111 points, Youden index 0.91; Low patience-distraction type, SOLPS total score ≥ 39 points, CII total score ≤ 88 points, Youden index 0.93.

[0096] S123. Preset the fitting hard threshold of the empirical calibration classification model for latent profile analysis, and calculate the classification accuracy of the model through the confusion matrix. Perform multiple rounds of accuracy verification on the pre-trained latent profile analysis model. When the accuracy meets the preset threshold, the solidified empirical calibration classification model is obtained.

[0097] Specifically, the preset hard thresholds for fitting include three items: 1. Entropy value ≥ 0.99. The closer the entropy value is to 1, the higher the clarity of the model classification and the smaller the classification error; 2. Lo-Mendell-Rubin likelihood ratio test (LMR) P value < 0.001; 3. Bootstrap likelihood ratio test (BLRT) P value < 0.001. Both tests with P values ​​< 0.001 indicate that the model fits significantly better than the model with fewer classes.

[0098] Precision verification employs a 10-fold cross-validation method, randomly dividing the training set into 10 groups: 9 groups for training and 1 group for validation. The confusion matrix is ​​calculated 10 times, and the average of classification accuracy, recall, precision, and F1 score is taken to ensure that the model achieves a classification accuracy ≥ 95%, recall ≥ 93%, and F1 score ≥ 94% on the validation set. Once all fitting hard thresholds and precision metrics meet the requirements, all model parameters and classification thresholds are locked, completing model solidification. The classification rules and threshold ranges of the solidified model will no longer change with each new batch of samples.

[0099] The score distribution of the 10 items of the SOLPS scale for the three potential categories of nursing students identified in this step is as follows: Figure 3 As shown in the figure, the SOLPS single-item scores of nursing students with low patience and distractibility remained stable in the range of 4.3-4.7, those with medium patience and fluctuating scores remained stable in the range of 2.9-3.1, and those with high patience and focus remained stable in the range of 1.4-1.6. The SOLPS item scores of the three types of nursing students showed completely separated inter-group differences, with intra-group score fluctuations not exceeding ±0.3 points. This demonstrates that the three optimal classification numbers determined in this invention have extremely strong discriminative power, the classification results exhibit high intra-group consistency and inter-group differences, and the corresponding classification threshold ranges have statistical rationality and practical application value.

[0100] S13. Based on the core anchoring feature dimension, auxiliary verification dimension and fixed weights, the nursing student assessment score is used as the output label. Generalized linear regression is pre-trained using an effective labeled dataset. Empirical calibration is performed by combining marginal contribution degree to obtain the empirical weight calibration prediction model.

[0101] In the description of this invention, the nursing student assessment score is used as the output label. A generalized linear regression pre-training process is performed using a valid labeled dataset, and empirical calibration is performed by combining marginal contribution. The resulting empirical weight calibration prediction model includes:

[0102] S131. Using the theoretical and practical assessment scores of nursing students as the output labels, select a valid labeled dataset as the pre-training sample, take the full set of standardized feature data in the pre-training sample as the input, and use the multiple linear regression algorithm to construct the initial prediction model and perform pre-training.

[0103] Specifically, the calculation method for the output label is as follows: Nursing student's final comprehensive assessment score = final theoretical closed-book assessment score × 60% + final nursing practical skills assessment score × 40%. Both scores are out of 100, and the comprehensive score ranges from 0 to 100, which is in line with the core goal of nursing professional talent training. The pre-training sample is completely consistent with the 764 samples, 7:1:2 division rule, and random seed used in S121, ensuring that the training benchmark of the classification model and the prediction model is consistent.

[0104] A multiple linear regression model with L2 regularization is used, and the model expression is as follows:

[0105] ;

[0106] Where Y represents the nursing student's final comprehensive assessment score. For the intercept term, Let be the regression coefficient of the i-th feature. For the i-th standardized feature, The random error term is; the L2 regularization term is... The regularization coefficient λ = 0.01.

[0107] The training set data is input into the initial model, and the batch gradient descent algorithm is used to optimize the model parameters. The loss function is the mean squared error (MSE). Before pre-training, the variance inflation factor (VIF) is used to test for multicollinearity among features, and collinear features with VIF>10 are removed to ensure that the Gauss-Markov assumptions of the multiple linear regression model are met.

[0108] S132. Combine the use of an interpretability algorithm (Tree-SHAP algorithm) to calculate the marginal contribution of each feature, and perform empirical calibration on the weight coefficients of the initial prediction model to calibrate the fixed weight coefficients of each feature.

[0109] Specifically, the Tree-SHAP algorithm is used to calculate the marginal contribution of each input feature to the prediction result of the comprehensive assessment score, and to verify whether the sign and magnitude of the regression coefficients of the initial model are consistent with the trend of the SHAP marginal contribution. For features whose regression coefficients are inconsistent with the trend of the SHAP contribution, they are removed by stepwise regression to avoid the features from causing reverse interference to the model results.

[0110] Finally, the regression coefficients of the initial model were normalized by combining the marginal contribution ratio of SHAP, and the final fixed weight coefficients of each feature were determined. Among them, the core feature weight of the online learning patience scale was 42%, the core feature weight of the exploration community scale was 33%, the total weight of the two core anchor feature dimensions was 75%, and the total weight of each feature of the auxiliary verification dimension was 25%.

[0111] This step uses SHAP marginal contribution to empirically calibrate the model weights, ensuring that the weight allocation is fully matched with the actual impact of features on the learning results. This avoids the weight bias caused by multicollinearity and outliers in conventional regression models, and significantly improves the accuracy of model predictions.

[0112] S133. Based on the lower limit of the confidence interval of the pre-training results, set a hard threshold for the goodness of fit of the generalized linear empirical weight calibration prediction model, and calculate the average absolute error between the model prediction value and the actual assessment score of nursing students. Verify the accuracy of the pre-trained model. When the goodness of fit and the average absolute error both meet the preset requirements, the solidified empirical weight calibration prediction model is obtained.

[0113] Specifically, the hard threshold for goodness of fit is a coefficient of determination R² ≥ 0.85. This threshold is determined based on the lower limit of the 95% confidence interval of the pre-training results. The closer R² is to 1, the stronger the model's ability to interpret the learning results of nursing students. The preset MAE ≤ 8% means that the average deviation between the model's predicted value and the actual assessment score does not exceed 8 points (out of 100).

[0114] Accuracy verification uses 10-fold cross-validation to ensure that the model has R² ≥ 0.85 and MAE ≤ 8% on the validation set. After all accuracy indicators meet the requirements, the model weights and all parameters are locked to complete model solidification. The solidified model weights will no longer be automatically adjusted with new samples.

[0115] S2. Perform two-dimensional hierarchical data collection based on the core anchoring feature dimension and the auxiliary verification dimension, and perform quality verification and deviation calibration on the collected data to output a standardized and effective feature set.

[0116] In the description of this invention, a two-dimensional hierarchical data acquisition is performed based on the core anchoring feature dimension and the auxiliary verification dimension. The acquired data undergoes quality verification and deviation calibration, and the output standardized effective feature set includes:

[0117] S21. Based on the core anchoring feature dimension and auxiliary verification dimension of the nursing student learning profile, perform dual-dimensional hierarchical data collection for the construction of the nursing student learning profile, simultaneously complete the basic quality automatic verification of the collected data, and generate the initial valid dataset for the construction of the nursing student learning profile after removing invalid data.

[0118] Specifically, the dual-dimensional hierarchical data collection is divided into initial baseline full-dimensional collection and subsequent dynamic verification collection. The initial baseline collection covers all fields of the core anchor feature dimension and the auxiliary verification dimension, while the subsequent dynamic verification collection only targets the core anchor feature dimension, and auxiliary verification dimension data is collected only when the full-dimensional update condition is triggered.

[0119] Core anchor feature dimension data is collected by pushing standardized packaged electronic questionnaires to nursing students through the nursing blended teaching platform. An electronic informed consent form is set up before the push, and nursing students can fill it out only after confirming their consent. After the completion, the platform automatically collects the original answer data and completes the structured coding. Auxiliary verification dimension data is automatically and synchronously collected through the connection with the school's academic affairs management system and the behavior tracking interface of the teaching platform. No manual operation is required, and the collection frequency is synchronized with the teaching cycle.

[0120] The rules for automated verification of basic quality include: automatic blocking of submissions with missing required fields; allowing only one valid baseline response per student's unique anonymized ID; marking responses to lie detection questions built into the scale as invalid if they do not meet preset requirements; marking responses with a total response time of less than 3 minutes as invalid; marking responses with consistent selection of the same option for all items as invalid; and marking samples with obvious logical contradictions in the responses as invalid.

[0121] S22. Based on the pre-trained empirical regression equation, identify and calibrate the self-assessment-behavioral contradiction samples in the initial effective dataset, perform standardized preprocessing on the verified samples, and output the standardized effective feature set for the fusion of nursing student learning profile features.

[0122] Specifically, the empirical regression equation for pre-training is the linear regression equation established during the S121 pre-training process, which relates the total SOLPS score to the video playback interruption rate and the timeliness of task completion. This equation is based on a large sample of 764 cases and clarifies the standard fitting relationship between nursing students' online learning patience score and the corresponding learning behavior data.

[0123] The identification rule for self-assessment-behavioral inconsistencies is as follows: calculate the deviation rate between the actual SOLPS score of nursing students and the theoretical score predicted by the regression equation, the deviation rate = |actual score - theoretical score| / theoretical score × 100%, the preset allowable threshold for deviation rate is ±15%, this threshold is determined based on the 95% prediction interval of the pre-training data, and samples with a deviation rate exceeding ±15% are marked as self-assessment-behavioral inconsistencies.

[0124] For contradictory samples, the system automatically pushes a scale review prompt. If the deviation rate still exceeds the standard after the nursing students complete the review, the SOLPS score is calibrated based on the theoretical score calculated by the regression equation from the behavioral data. The calibration coefficient is determined entirely based on the pre-trained regression coefficient, with no room for subjective adjustment.

[0125] Standardization preprocessing includes: performing min-max standardization on numerical data to uniformly map it to the [0,1] interval; performing one-hot encoding on categorical variables such as gender, school system, and grade to convert them into computable numerical features; and filling missing values ​​with the median of the same latent category and dimension features calibrated by S12 to avoid population bias caused by mean filling.

[0126] S3. Using the fixed weights that have been labeled, perform weighted deep fusion on the standardized effective feature set to generate a fused feature vector corresponding to a single nursing student. Then, input the fused feature vector into the empirical labeling classification model and the empirical weight labeling prediction model to generate and store the initial nursing student learning profile.

[0127] In the description of this invention, as Figure 2As shown, a weighted deep fusion of standardized effective feature sets is performed using pre-defined fixed weights to generate a fused feature vector corresponding to a single nursing student. This fused feature vector is then input into the empirically calibrated classification model and the empirically weighted prediction model to generate and store the initial nursing student learning profile, including:

[0128] S31. Perform weighted feature fusion on the standardized effective feature set according to the fixed weights, and generate a unique one-dimensional fusion feature vector corresponding to each nursing student.

[0129] Specifically, the calculation formula for weighted feature fusion is: Fusion feature vector = Σ (standardized feature value of a single nursing student × pre-calibrated fixed weight of the corresponding feature), where the weight of each feature is the fixed weight calibrated in S113, and the sum of the weights is 1; the fusion process strictly follows the pre-calibrated weights and does not adjust with the feature data of a single nursing student, ensuring that the generation rules of the fusion feature vector of all nursing students are completely unified; the final generated one-dimensional fusion feature vector is a single-value vector with a value range of [0,1]. The higher the vector value, the better the nursing student's comprehensive learning status and the higher the expected learning result.

[0130] This step is based on weighted fusion with fixed weights determined by empirical calibration, which ensures that the fused feature vector always takes the core anchor features as the core, avoiding the problem of core features being submerged due to equal weight fusion and automatic weight learning. The Pearson correlation coefficient between the fused feature vector and the actual learning results of nursing students is ≥0.7, which is much higher than that of conventional fusion methods.

[0131] S32. Calculate the Pearson correlation coefficient between a single fused feature vector and the nursing student's historical learning results, and compare it with the preset correlation coefficient threshold. Automatically return samples that fail the verification to re-execute data collection, quality verification and deviation calibration.

[0132] Specifically, the Pearson correlation coefficient is used to measure the linear correlation between the fused feature vector and the average score of the nursing student's last three assessments. The correlation coefficient ranges from -1 to 1, and the closer the absolute value is to 1, the stronger the correlation.

[0133] The preset correlation coefficient threshold is ≥0.7. This threshold is determined based on the lower limit of the 95% confidence interval of the correlation coefficient between the fused feature vector and the learning result in the pre-training data. A correlation coefficient ≥0.7 is considered to have passed the verification and can proceed to the subsequent classification and prediction stages. A correlation coefficient <0.7 is considered to have failed the verification, indicating that the fused feature vector cannot effectively reflect the actual learning status of the nursing students. The system will automatically push a data review notification to the nursing students and re-execute the core anchor dimension data collection, quality verification and deviation calibration process.

[0134] The feature fusion validity verification step in this process prevents invalid data from entering subsequent model calculations from the source, ensuring that the input data for profile construction is strongly correlated with the learning results of nursing students, thus fundamentally guaranteeing the accuracy of the profile.

[0135] S33. Input the validated fused feature vector into the empirically calibrated classification model, and output the potential category affiliation and classification confidence for each nursing student. Then input the fused feature vector into the empirical weight calibration prediction model, and output the predicted value of the nursing student's learning result. Based on the predicted value, the learning risk level is calibrated.

[0136] Specifically, the empirically calibrated classification model is the S12-based latent profile analysis empirically calibrated classification model. The model automatically calculates the posterior probabilities of three latent categories corresponding to a single nursing student's fusion feature vector. The category with the highest posterior probability is determined as the latent category to which the nursing student belongs, and the classification confidence score, i.e., the highest posterior probability value, is output. A classification confidence score ≥ 95% is preset as a valid classification result; samples with a confidence score below 95% are automatically subject to manual review.

[0137] The empirical weight calibration prediction model is the generalized linear empirical weight calibration prediction model fixed in S13. The model automatically outputs the predicted value of the nursing students' final comprehensive assessment score, and the predicted value is on a percentage scale. The learning risk level based on the predicted value is: predicted value <60 points is high risk (extremely high risk of failing the course), 60-70 points is medium risk (poor learning status), and ≥70 points is low risk (stable learning status).

[0138] S34. Map the potential category affiliation, predicted outcome values, and standardized feature data of nursing students into multi-layered structured labels, and set a quantitative value range for each label. Generate an initial nursing student learning profile for each nursing student based on the full-dimensional labels, and store it in the profile database after associating it with the nursing student's anonymous ID.

[0139] Specifically, the multi-layered structured label consists of six layers: core anchoring label, auxiliary verification label, behavioral feature label, learning experience label, ability prediction label, and risk warning label. Each label has a clearly defined quantitative value range and no vague qualitative descriptions.

[0140] The core anchoring labels include three quantitative fields: SOLPS score, CII score, and potential category affiliation; the auxiliary verification labels include two quantitative fields: basic attributes and historical assessment scores; the behavioral characteristic labels include three quantitative fields: video playback interruption rate, task completion timeliness, and interactive participation; the learning experience labels include four quantitative fields: scores for each of the four dimensions of the CII scale; the ability prediction label includes one quantitative field: predicted final exam score; and the risk warning label includes one quantitative field: risk level. The initial nursing student learning profile is the baseline profile before the start of the nursing student course, containing all the quantitative data of the above six-layer labels across all dimensions. After the profile data is associated with the nursing student's unique anonymized ID, it is encrypted and stored in the teaching platform's dedicated profile database, accessible only to authorized instructors and teaching administrators according to their permissions.

[0141] S4. Based on the core anchor feature dimension, a time- and event-driven dynamic update mechanism is constructed. When the update threshold is reached, the nursing student learning profile is updated, and the profile change trajectory is retained.

[0142] In the description of this invention, a time- and event-driven dynamic update mechanism is constructed based on the core anchoring feature dimension. When the update threshold is reached, the update of the nursing student learning profile is triggered, and the profile change trajectory is preserved, including:

[0143] S41. Based on the longitudinal tracking data of the actual collected nursing student samples, the natural fluctuation range of nursing students' online learning patience is statistically obtained, and the change of the online learning patience scale score is selected as a preset threshold as the substantive change threshold of the core anchor feature dimension.

[0144] Specifically, the longitudinal tracking data of the actual nursing student sample consisted of 764 nursing students who underwent three SOLPS scale retests over a 16-week teaching period. The statistical analysis of this longitudinal data showed that the natural fluctuation range of nursing students' patience in online learning was ±12% (95% confidence interval). In other words, under the condition that there is no substantial change in the learning status, the normal fluctuation of the SOLPS score will not exceed ±12%.

[0145] Pearson correlation analysis confirmed that when the change in SOLPS score was ≥20%, the correlation coefficient with the subsequent decline in nursing students' learning outcomes reached 0.78, which was significantly strong, indicating that the learning status of nursing students had undergone substantial changes.

[0146] Based on ROC curve analysis, when the threshold is set at 20%, the sensitivity for identifying substantial changes in learning status reaches 92.3%, the specificity reaches 94.7%, and the Youden index is the highest. Therefore, the threshold for substantial changes in the core anchor feature dimension is selected as ≥20%. That is, only when the combined score of SOLPS and CII changes by ≥20% is it determined that there has been a substantial change in the nursing student's learning status. This threshold is the only necessary condition for the full-dimensional update of the nursing student's learning profile. The reasonable adjustment range of the threshold is 15%-25%, which can be adapted according to the teaching cycle of different institutions.

[0147] S42. Set time-driven anchor point review updates to match the blended teaching cycle of nursing students. Preset multiple fixed anchor point review nodes. When a fixed node is reached, push the online learning patience scale and the inquiry community scale. When the change in the retest score is greater than or equal to the substantial change threshold, trigger the full collection of auxiliary verification dimension data and the full-dimensional update of nursing student learning profile.

[0148] Specifically, matching the 16-week / semester standard blended learning cycle of the nursing major, three fixed anchor point review nodes are preset: a core anchor dimension review once in the 8th week of the middle of the course, a full dimension review in the 16th week of the end of the course, and a baseline retest review in the first week of the new semester.

[0149] When the fixed review node is reached, the teaching platform only pushes the SOLPS and CII core anchor scales to nursing students and does not collect other auxiliary dimension data. Only when the change of the nursing student's retest score and the course baseline score is ≥20%, the full collection of auxiliary verification dimension data, feature fusion, and full-dimensional update of the nursing student learning profile are automatically triggered. If the change of the retest score is <20%, only the time node record of the nursing student learning profile is updated, and the full-dimensional reconstruction is not triggered. The core classification and prediction results of the profile remain unchanged.

[0150] S43. Set up event-driven anchor point linkage updates, preset multiple types of pre-triggered conditions. When any pre-triggered event is triggered, the core anchor dimension retest is executed first. When the retest score is greater than or equal to the substantial change threshold, the full-dimensional data collection and nursing student learning profile update are triggered.

[0151] Specifically, four types of pre-triggered events are preset. All events are based on the verification of pre-training results and are strongly correlated with changes in the core anchoring dimensions of nursing students. These events are: 1. Nursing students actively submit applications for retesting the SOLPS scale; 2. The interruption rate of course videos is ≥85% for four consecutive sessions and assignments are overdue for more than two consecutive sessions; 3. The score of a single unit assessment drops by ≥35 points compared to the previous one; 4. The instructor manually triggers the anchor point review based on the nursing students' classroom performance.

[0152] When any pre-event is triggered, the teaching platform immediately pushes the core anchor scale to nursing students to complete the retest. Only when the change in the retest score from the baseline score is ≥20% will the full-dimensional data collection, feature fusion, and full-dimensional update of the nursing student learning profile be triggered. If the change in the retest score is <20%, only the corresponding behavioral tags will be updated, without changing the core classification and prediction results of the profile.

[0153] S44. When the number of newly added valid labeled samples in the portrait database reaches the initial pre-training sample size, the model retraining process is triggered, and the newly added valid samples are added to the training set to re-optimize the model parameters of the empirical labeling classification model and the empirical weight labeling prediction model.

[0154] Specifically, the initial pre-training sample size is 764 cases. The preset retraining trigger condition is: the number of newly added valid labeled samples in the portrait database reaches 50% of the initial pre-training sample size, i.e., ≥382 new valid labeled samples. After triggering retraining, the newly added valid samples are merged with the initial pre-training samples to form a new training set, and the parameters of the latent profile analysis empirical calibration classification model and the generalized linear empirical weight calibration prediction model are refitted.

[0155] After model retraining, the fitting threshold verification and 10-fold cross-validation must be performed in the same manner as S12 and S13. Model replacement is only performed if the classification accuracy and goodness of fit of the model after iteration are not lower than the original calibration model, and the deviation of the proportion of the three potential categories of nursing students does not exceed ±5%. If the accuracy of the model decreases after iteration, the original calibration model is retained and replacement is not performed.

[0156] S45. Synchronously store the dynamically updated portrait data and model iteration records in the portrait database, and retain the portrait change trajectory throughout the entire learning cycle of nursing students.

[0157] Specifically, each dynamically updated profile data, including update time, core anchor dimension score, potential category attribution, prediction result, risk level, and full-dimensional label data, is synchronously stored in the profile database along with the nursing student's anonymized ID.

[0158] Model iteration records, including retraining time, new sample size, model parameter changes, and accuracy verification results, are encrypted and stored in the model management database. A unique profile change trajectory file is created for each nursing student, fully recording the core data changes of each profile update throughout the entire learning cycle from the start to the end of the course, supporting the retrospective tracking of the entire learning status changes of nursing students by time. The storage period of profile data and trajectory files strictly follows the relevant regulations for educational data management, and automatic anonymization and destruction are performed 3 years after the nursing student graduates.

[0159] S5. Based on the potential category attribution of the dynamically updated nursing student learning profile, match the corresponding intervention push and risk warning, collect intervention effect feedback data, and perform reverse verification on the fixed weights.

[0160] In the description of this invention, based on the potential category attribution of the dynamically updated nursing student learning profile, corresponding intervention push and risk warning are matched, and feedback data on intervention effects are collected. Reverse verification of fixed weights includes:

[0161] S51, a library of empirically matched intervention strategies that pre-determine one-to-one correspondences with the three potential categories of nursing students.

[0162] Specifically, for nursing students with high patience and focus, the strategies include: matching extended learning tasks, guiding research capabilities, and peer teaching interventions; for nursing students with medium patience and fluctuating levels, the strategies include: learning task decomposition, real-time learning feedback, group collaborative learning, and structured reflection interventions; and for nursing students with low patience and distractibility, the strategies include: low cognitive load instructional design, one-on-one learning guidance, emotional support, and focus training interventions. Each intervention strategy has a clearly defined delivery method, execution cycle, and implementation standards.

[0163] Specifically, all strategies in the intervention strategy library are based on empirical research findings among nursing students, and have been validated by a large sample with an effectiveness rate of ≥60% for the corresponding categories of nursing students, rather than being subjectively designed based on expert experience. For intervention strategies targeting highly patient and focused nursing students, the delivery method involves targeted pushes of extended learning resources, introductory research courses, and peer teaching volunteer recruitment notices to the teaching platform, implemented once every two weeks, with an implementation standard of ≥60% student participation in completing extended tasks. For moderately patient and fluctuating nursing students, the delivery method involves instructors receiving strategy suggestions and implementing them in their teaching design, while the platform pushes a breakdown of learning tasks and daily learning reminders to students, implemented once a week, with an implementation standard of over 80% on-time task completion rate. For low-patient and distracted nursing students, the delivery method involves one-on-one communication between counselors and instructors, and the platform pushing low-cognitive-load fragmented learning resources and focus training mini-programs, implemented twice a week.

[0164] S52. Based on the potential category classification of nursing students' baseline and dynamic profiles, push matching intervention strategies to the corresponding instructors. For nursing students with high risk levels, automatically push early warning information containing risk causes and empirical intervention suggestions to instructors and counselors.

[0165] Specifically, the teaching platform automatically synchronizes updated profile data for all nursing students weekly. Based on the students' current potential category, it pushes matching intervention strategy suggestions to the instructors of the corresponding courses. Instructors can view strategy details and the corresponding list of nursing students on the teaching platform and apply them to their teaching designs with one click. For high-risk nursing students, the system automatically triggers a red alert and pushes warning information to instructors and class counselors in real time. The information includes the student's risk level, the core risk cause, and the corresponding empirical intervention suggestions. It also supports instructors and counselors to record the implementation of interventions online. The warning information is pushed once a week for high-risk nursing students, once every two weeks for medium-risk nursing students, and no warning information is pushed to low-risk nursing students.

[0166] S53. After the intervention strategy is implemented, the system automatically collects feedback data on changes in nursing students' online learning patience scale scores, exploration community scale scores, online learning behavior improvement, and subsequent assessment results. The preset effective intervention standard is the improvement in nursing students' core anchoring characteristic dimension scores after the intervention.

[0167] Specifically, the intervention strategy is implemented over a period of 4 weeks. After implementation, the teaching platform automatically collects feedback data on the intervention effect, including changes in SOLPS and CII scale scores before and after the intervention, changes in behavioral data such as video playback interruption rate, task completion rate, and interactive participation, as well as changes in unit assessment scores. The preset effective intervention criteria are to meet any of the following conditions: 1. The improvement in nursing students' SOLPS and CII core anchor dimensions scores after the intervention is ≥15%; 2. The improvement in nursing students' unit assessment scores after the intervention is ≥10% compared to before the intervention.

[0168] At the end of each semester, the system automatically calculates the overall effectiveness of each type of intervention strategy. Effectiveness = number of nursing students who received effective intervention / total number of nursing students who received intervention in that category × 100%. When the overall effectiveness of a certain type of intervention strategy is lower than 60%, the system automatically triggers strategy optimization. Based on newly added effective intervention sample data, the system uses correlation analysis to screen measures that are strongly correlated with the intervention effect, replaces measures with low effectiveness in the original strategy library, and regenerates the corresponding type of intervention strategy.

[0169] S54. Based on the feedback data of the intervention effect, the feature system, contradiction identification rules and feature fusion weights are verified in reverse. Only when the feedback data shows that the change in the SHAP contribution of a certain feature to the learning effect of nursing students is ≥0.2, the weights and model parameters of the corresponding dimension are adjusted. Otherwise, the original calibration standards are kept unchanged.

[0170] Specifically, at the end of each academic year, based on the full intervention effect feedback data and nursing student learning outcome data, the Tree-SHAP algorithm is recalculated to determine the marginal contribution of all features to the nursing student learning outcomes, and compared with the initial SHAP contribution value calibrated in S11. Only when the absolute value of the SHAP contribution of a feature changes by ≥0.2 is it determined that the influence of that feature on the learning outcome has changed substantially, triggering the corresponding dimension's weight adjustment, contradiction identification rule optimization, and model parameter retraining. If the change in the feature's SHAP contribution is <0.2, the original calibrated feature system, weight coefficients, and model parameters remain unchanged. Weight adjustment and model retraining must strictly follow the calibration process and verification standards of S11-S13 to ensure that the accuracy of the adjusted model is not lower than that of the original model.

[0171] Please see Figure 4 It also provides a dynamic construction system for nursing student learning profiles based on multi-source data fusion, which includes:

[0172] Baseline model construction module 1 is used to perform feature selection on nursing student learning profiles based on the nursing student full-dimensional feature dataset, select the core anchor feature dimensions and auxiliary verification dimensions of nursing student learning profiles, and empirically calibrate the fixed weights of each dimension in order to construct an empirical calibration classification model and an empirical weight calibration prediction model.

[0173] The data acquisition and identification module 2 is used to perform two-dimensional hierarchical data acquisition based on the core anchor feature dimension and the auxiliary verification dimension, and to perform quality verification and deviation calibration on the acquired data, and output a standardized and effective feature set.

[0174] The portrait fusion generation module 3 is used to perform weighted deep fusion on the standardized effective feature set using the calibrated fixed weights to generate a fusion feature vector corresponding to a single nursing student. The fusion feature vector is then input into the empirical calibration classification model and the empirical weight calibration prediction model to generate and store the initial nursing student learning portrait.

[0175] The dynamic iterative update module 4 is used to build a time- and event-driven dynamic update mechanism based on the core anchor feature dimension. When the update threshold is reached, the update of the nursing student learning profile is triggered, and the profile change trajectory is retained.

[0176] The intervention closed-loop verification module 5 is used to match the corresponding intervention push and risk warning based on the potential category classification of the dynamically updated nursing student learning profile, and to collect intervention effect feedback data to perform reverse verification on the fixed weight.

[0177] In summary, by employing the technical solution described above, and through a combination of fully relevant feature screening and interpretability analysis, the core dimensions of the nursing student learning profile are accurately identified and their corresponding weights are empirically calibrated. This overcomes the limitations of conventional empirical feature selection, establishing a strong correlation between the core dimensions of the profile and the nursing student's learning outcomes. Simultaneously, the fixed weight design ensures the consistency of the profile benchmark, avoiding the problem of core features being overwhelmed by noise due to automatic weight optimization by the algorithm, effectively improving the profile's group adaptability and result stability. By constructing a classification and prediction model adapted to the characteristics of the nursing student group, combined with a dual-mode dynamic update mechanism triggered by core anchor points, the core problems of insufficient model generalization and redundant update operations in conventional techniques are solved. This ensures that the profile can accurately capture substantial changes in the nursing student's learning state while avoiding the system resource consumption caused by meaningless full updates, achieving a two-way balance between profile recognition accuracy and system operating efficiency. By identifying and calibrating the contradictions between self-assessment data and behavioral data, the whole-process quality control system for subjective data has been improved, effectively enhancing the effectiveness of the profile input data. At the same time, by combining the empirical intervention matching of profile classification results and the reverse optimization design of intervention effect feedback, a complete technical closed loop from profile construction to teaching implementation has been formed, making the profile technology truly adaptable to the personalized training needs of blended nursing education and greatly enhancing the practical application value of the technology.

[0178] This application also provides an electronic device, such as... Figure 5 As shown, it includes: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device executes the method and system for dynamically constructing nursing student learning profiles based on multi-source data fusion as described in any of the above embodiments.

[0179] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0180] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0181] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0182] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0183] This application also provides a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a flexible component distribution medium, etc.

[0184] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

Claims

1. A method for dynamically constructing nursing student learning profiles based on multi-source data fusion, characterized in that, include: Based on the full-dimensional feature dataset of nursing students, feature selection was performed on the nursing student learning profile. The core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile were selected, and the fixed weights of each dimension were empirically calibrated in order to construct an empirical calibration classification model and an empirical weight calibration prediction model. Data collection is performed in a two-dimensional hierarchical manner based on the core anchoring feature dimension and the auxiliary verification dimension. The collected data is then subjected to quality verification and deviation calibration to output a standardized and effective feature set. By using the fixed weights of the standardization effective feature set, a weighted deep fusion is performed to generate a fused feature vector corresponding to a single nursing student. The fused feature vector is then input into the empirical calibration classification model and the empirical weight calibration prediction model to generate and store the initial nursing student learning profile. Based on the core anchor feature dimension, a time- and event-driven dynamic update mechanism is constructed. When the update threshold is reached, the nursing student learning profile is updated, and the profile change trajectory is preserved. Based on the potential category attribution of the dynamically updated nursing student learning profile, corresponding intervention push and risk warning are matched, and feedback data on intervention effect are collected to perform reverse verification on fixed weights.

2. The method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to claim 1, characterized in that, The process of feature selection for nursing student learning profiles based on a comprehensive dataset of nursing student features, selecting core anchoring feature dimensions and auxiliary verification dimensions, and empirically calibrating the fixed weights of each dimension to construct an empirically calibrated classification model and an empirically calibrated prediction model includes: We used a fully relevant feature selection algorithm combined with interpretability analysis to select effective features from the full-dimensional feature dataset of nursing students and quantified the marginal contribution of each effective feature to the prediction of nursing students' learning outcomes. Based on the marginal contribution results, we divided the core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile and set fixed weights for each dimension. Based on the core anchoring feature dimension, auxiliary verification dimension and fixed weight, the potential profile analysis pre-training is carried out using the nursing student effective labeled dataset, and the accuracy is verified by the classification threshold. After the verification is passed, it is solidified into the empirical labeling classification model of potential profile analysis. Based on the core anchoring feature dimension, auxiliary verification dimension, and fixed weights, the nursing student assessment score is used as the output label. Generalized linear regression is pre-trained using an effective labeled dataset, and empirical calibration is performed by combining marginal contribution degree to obtain an empirical weight calibration prediction model.

3. The method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to claim 2, characterized in that, The method employs a fully correlated feature screening algorithm combined with interpretability analysis to screen effective features in the full-dimensional feature dataset of nursing students and quantifies the marginal contribution of each effective feature to the prediction of nursing students' learning outcomes. Based on the marginal contribution results, the core anchoring feature dimension and auxiliary verification dimension of the nursing student learning profile are divided, and fixed weights are set for each dimension, including: Shaded features with random permutations are created from the original features in the full-dimensional feature dataset of nursing students, and the feature importance of the original features and the shaded features is verified by two-sided statistical analysis. Based on a preset significance threshold, all effective features that are statistically correlated with the learning results of nursing students are iteratively selected. An interpretability algorithm was used to calculate the marginal contribution of all effective features to the prediction of nursing students' learning outcomes. The features were sorted from high to low according to their marginal contribution. The online learning patience scale and the exploration community scale, which had the highest comprehensive interpretability contribution, were selected as the core anchor feature dimensions for the nursing students' learning profile. The remaining effective features were divided into auxiliary verification dimensions according to their contribution attributes. Based on the interpretability contribution ratio of each feature, the fixed weights of each feature within the core anchor feature dimension and the auxiliary verification dimension are empirically determined. The core anchoring feature dimensions were digitized and standardized, and the tiered scoring rules, total score range, and reverse scoring conversion rules for the online learning patience scale were selected. The tiered scoring rules, total score range, and positive scoring conversion rules for the exploration community scale were also selected.

4. The method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to claim 2, characterized in that, The process of pre-training for latent profile analysis using a nursing student-labeled dataset and verifying accuracy through a classification threshold, followed by solidification into an empirically labeled classification model for latent profile analysis after passing the verification, includes: Several sets of effective labeled datasets of nursing students collected in practice were selected as pre-training samples. The standardized feature data of the core anchoring feature dimension and auxiliary verification dimension in the pre-training samples were used as input and substituted into the potential profile analysis model for pre-training. The model fitting effect of different numbers of categories was compared one by one by the likelihood ratio test to select the optimal number of categories. The feature discrimination of each category of nursing students was analyzed by the receiver operating characteristic curve. Combined with the Youden index calculation results, the optimal classification threshold range of each category of nursing students was selected. A hard threshold for fitting the empirically calibrated classification model for latent profile analysis is preset, and the classification accuracy of the model is calculated through the confusion matrix. The pre-trained latent profile analysis model is subjected to multiple rounds of accuracy verification. When the accuracy meets the preset threshold, the solidified empirically calibrated classification model is obtained.

5. The method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to claim 2, characterized in that, The empirical weighting prediction model, which uses nursing students' assessment scores as output labels, is pre-trained using a generalized linear regression model based on an effectively labeled dataset, and then empirically calibrated using marginal contribution, includes: The assessment scores calculated from the theoretical and practical assessment scores of nursing students are used as output labels. Valid labeled datasets are selected as pre-training samples. All standardized feature data in the pre-training samples are used as input. The initial prediction model is constructed and pre-trained using the multiple linear regression algorithm. By combining the calculation of the marginal contribution of each feature using an interpretable algorithm, the weight coefficients of the initial prediction model are empirically calibrated to calibrate the fixed weight coefficients of each feature. Based on the lower limit of the confidence interval of the pre-training results, a hard threshold for the goodness of fit of the generalized linear empirical weight calibration prediction model is set, and the mean absolute error between the model prediction value and the actual assessment score of nursing students is calculated. The accuracy of the pre-trained model is verified. When both the goodness of fit and the mean absolute error meet the preset requirements, the solidified empirical weight calibration prediction model is obtained.

6. The method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to claim 1, characterized in that, The process of using pre-defined fixed weights to perform weighted deep fusion on the standardized effective feature set to generate a fused feature vector corresponding to a single nursing student, and then inputting the fused feature vector into the empirically calibrated classification model and the empirically weighted prediction model to generate and store the initial nursing student learning profile, includes: The standardized effective feature set is weighted and fused according to the fixed weights to generate a unique one-dimensional fused feature vector corresponding to each nursing student. Calculate the Pearson correlation coefficient between a single fused feature vector and the nursing student's historical learning results, and compare it with the preset correlation coefficient threshold. Samples that fail the verification are automatically returned to re-execute data collection, quality verification and deviation calibration. The validated fusion feature vector is input into the empirical calibration classification model, which outputs the potential category affiliation and classification confidence of a single nursing student; the fusion feature vector is input into the empirical weight calibration prediction model, which outputs the predicted value of the nursing student's learning result, and the learning risk level is calibrated based on the predicted value. The potential category affiliation, predicted outcome value, and standardized feature data of nursing students are mapped into multi-layer structured labels, and a quantitative value range is set for each label. An initial nursing student learning profile is generated based on the full-dimensional labels, and then stored in the profile database after being associated with the nursing student's anonymous ID.

7. The method for dynamically constructing nursing student learning profiles based on multi-source data fusion according to claim 1, characterized in that, The aforementioned dynamic update mechanism, based on core anchored feature dimensions and driven by time and events, triggers an update of the nursing student learning profile when an update threshold is reached, and retains the profile's change trajectory, including: Based on longitudinal tracking data of actual nursing student samples, the natural fluctuation range of nursing students' patience in online learning was statistically obtained, and the change of the online learning patience scale score greater than or equal to the preset threshold was selected as the substantive change threshold of the core anchor feature dimension. Set up time-driven anchor point review updates to match the blended teaching cycle of nursing students. Preset multiple fixed anchor point review nodes. When a fixed node is reached, push online learning patience scale and inquiry community scale. When the change in retest score is greater than or equal to the substantial change threshold, trigger the full collection of auxiliary verification dimension data and full-dimensional update of nursing student learning profile. Set up event-driven anchor point linkage updates, preset multiple types of pre-trigger conditions. When any pre-trigger event is triggered, the core anchor dimension retest is executed first. When the retest score is greater than or equal to the substantial change threshold, the full-dimensional data collection and nursing student learning profile update are triggered. When the number of newly added valid labeled samples in the portrait database reaches the initial pre-training sample size, the model retraining process is triggered, and the newly added valid samples are added to the training set to re-optimize the model parameters of the empirical labeling classification model and the empirical weight labeling prediction model. Each dynamically updated profile data and model iteration record are synchronously stored in the profile database, and the profile change trajectory throughout the entire learning cycle of nursing students is preserved.

8. A dynamic construction system for nursing student learning profiles based on multi-source data fusion, used to implement the dynamic construction method for nursing student learning profiles based on multi-source data fusion as described in any one of claims 1-7, characterized in that, The system includes: The baseline model building module is used to perform feature selection on the nursing student learning profile based on the nursing student full-dimensional feature dataset, select the core anchor feature dimension and auxiliary verification dimension of the nursing student learning profile, and perform empirical calibration on the fixed weight of each dimension to build an empirical calibration classification model and an empirical weight calibration prediction model. The data acquisition and recognition module is used to perform two-dimensional hierarchical data acquisition based on the core anchor feature dimension and the auxiliary verification dimension, and to perform quality verification and deviation calibration on the acquired data, and output a standardized and effective feature set. The profile fusion generation module is used to perform weighted deep fusion on the standardized effective feature set using the labeled fixed weights, generate a fusion feature vector corresponding to a single nursing student, and input the fusion feature vector into the empirical labeling classification model and the empirical weight labeling prediction model to generate and store the initial nursing student learning profile. The dynamic iterative update module is used to build a time- and event-driven dynamic update mechanism based on the core anchor feature dimension. When the update threshold is reached, the nursing student learning profile is updated and the profile change trajectory is preserved. The intervention closed-loop verification module is used to match the corresponding intervention push and risk warning based on the potential category classification of the dynamically updated nursing student learning profile, and to collect intervention effect feedback data to perform reverse verification on fixed weights.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the method for dynamically constructing nursing student learning profiles based on multi-source data fusion as described in any one of claims 1 to 7 when executing a computer program.

10. A computer-readable storage medium, characterized in that, A computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the method for dynamically constructing nursing student learning profiles based on multi-source data fusion as claimed in any one of claims 1 to 7.

Images