An intelligent prediction system for depression change trajectory of college students based on multi-modal data fusion

Abstract

The application provides a college student depression change trajectory intelligent prediction system based on multi-modal data fusion, comprising: a multi-source data fusion acquisition module, a label construction unit, a model construction and selection module, a real-time prediction unit, a result output unit, a Web application module and a privacy protection and compliance unit. Through multi-source data fusion and a three-level quality control mechanism, the data efficiency is improved to more than 92%, which not only avoids the deviation problem of single data source, but also relies on longitudinal data collection for 4 years to capture dynamic changes, solves the limitations of traditional static evaluation, and provides comprehensive and reliable data support for mental health prediction.

Description

Technical Field

[0001] This invention relates to the technical field of trajectory prediction, and more particularly to an intelligent prediction system for the trajectory of changes in depression in college students based on multimodal data fusion. Background Technology

[0002] With increasing social competition and a faster pace of life, the mental health of university students has become a key focus of public health, with the incidence of depression showing a year-on-year upward trend. Depression not only causes core symptoms such as persistent low mood and loss of interest in university students, but also significantly affects their cognitive function, academic performance, and social adaptability, and in severe cases, can even lead to extreme events. Therefore, accurately characterizing the dynamic trajectory of depressive states in university students and effectively predicting potential risks of transformation are core prerequisites for building an early intervention system and improving the mental health of the population, and have significant public health value and educational practical significance. However, current technical systems in the field of predicting depression in university students still have many methodological limitations and are difficult to meet practical application needs.

[0003] The primary limitation of existing depression prediction technologies lies in their reliance on a single data modality, failing to construct a comprehensive model representing depressive states across multiple dimensions: physiological, psychological, and social. On one hand, some studies rely solely on psychological assessment scales (such as the BDI 2nd edition and SCL-90) for prediction. This data is essentially subjective reporting information, easily influenced by social expectation bias and attitudes towards responses (such as perfunctory answers or concealment of negative emotions), leading to compromised reliability and validity. More critically, single questionnaire data cannot reveal the physiological mechanisms behind changes in depressive states. For example, psychological scales cannot identify subtle abnormalities in brain structure and function (such as changes in cortical thickness or abnormal functional connectivity in brain regions), making it difficult to capture early physiological signals of potential shifts in depression. This results in insufficient early warning capabilities for the transition from the asymptomatic to the symptomatic phase in predictive models. On the other hand, other studies focus on the collection of single-modality physiological data (such as resting-state EEG, resting-state near-infrared spectroscopy, and structural magnetic resonance imaging data). Although these studies can reflect the association between brain activity or structural characteristics and depression from a neurophysiological perspective, they lack integrated analysis of basic information of college students (such as age and gender), demographic variables (such as family economic status and parenting style), and dynamic data on psychological state. This makes it impossible to establish a correlation chain of "changes in physiological indicators - evolution of psychological state - influence of social environment," which makes it difficult for the model to accurately interpret the pathological significance of depression behind physiological signals, ultimately limiting the comprehensiveness and accuracy of prediction.

[0004] Another core deficiency of existing technologies lies in the static nature of the assessment paradigm, which fails to capture the dynamic evolution of depressive states in college students. College students' depressive states are influenced by dynamic factors such as academic pressure (e.g., exams, thesis defenses), social interactions (e.g., interpersonal conflicts), and career development (e.g., internships, job hunting), exhibiting significant stage-specific fluctuations at key points such as the adjustment period upon entering university, critical academic periods, and the graduation transition. However, most current assessments still employ a "single cross-sectional survey" model (e.g., annual mental health surveys conducted by universities), which cannot obtain continuous data on the evolution of depressive states over time, nor can it identify the temporal patterns of potential transitions such as "mild depression to moderate depression" or "depression remission to relapse." This static assessment paradigm makes it easy to miss early risk signals. For example, if a student experiences a rapid deterioration in depressive symptoms due to academic failure within the interval between two surveys, existing technologies cannot capture this dynamic change in time, thus missing the optimal intervention window and exacerbating the risk of disease progression.

[0005] Existing technologies have significant shortcomings in data quality control mechanisms, further impacting the reliability of predictive models. During data collection, existing studies often lack standardized quality screening processes, leading to the inclusion of low-quality data (such as logically contradictory data and invalid responses) in the analysis sample. Specifically, some questionnaire data exhibit issues such as logical inconsistencies in items (e.g., high scores on the BDI-II "self-blame" dimension coexisting with low scores on the SCL-90 "hostility" and "interpersonal sensitivity" factors), and abnormal response times (e.g., completing a 90-item SCL-90 scale in less than 8 minutes). Physiological data may suffer from noise interference due to insufficient participant cooperation (e.g., frequent eye movements and head shaking during resting-state EEG collection) and equipment calibration deviations. Due to the lack of multi-dimensional data consistency verification rules (e.g., internal questionnaire reliability testing and physiological data signal-to-noise ratio screening), this low-quality data is not effectively removed, directly interfering with the subsequent feature extraction of depressive trajectories and the identification of potential turning points, resulting in systematic biases in model prediction results.

[0006] At the methodological level of model construction and validation, existing technologies also have significant limitations. First, the choice of model algorithms is often too simplistic. Most studies employ only traditional statistical methods (such as simple logistic regression) or a single machine learning algorithm (such as the basic random forest model), failing to fully leverage the advantages of multiple algorithms for comparative optimization. For example, they do not incorporate latent class growth models to characterize depression trajectory subtypes, nor do they use latent transition analysis (LTA) to identify the transition probabilities between different depression subtypes. This results in the model failing to fully explore the nonlinear features and interaction effects related to the dynamic changes in depression within the data. Second, the model validation process is not rigorous. Most studies only use internal cross-validation (such as 10-fold cross-validation), lacking independent, large-scale external sample validation (such as multi-center university samples), and failing to evaluate the model's generalization ability across different population subgroups (such as students from different majors or different places of origin). For example, some universities develop their own models based on data from only 500 students in their own school for training and validation. When these models are extended to other universities, the heterogeneity of the samples causes the prediction accuracy to drop from 85% in the training set to less than 60% in actual applications, making it difficult to meet the needs of cross-scenario applications.

[0007] Furthermore, existing technologies generally suffer from barriers in human-computer interaction and result transformation, limiting the practical application value of prediction results. Most current depression prediction systems only possess basic data storage and simple result output functions, lacking user-oriented visual interactive design: on the one hand, the result presentation format is simplistic (e.g., pure data tables), failing to visually demonstrate the changing trends and potential risks of depression through trajectory line graphs, risk heat maps, subtype radar charts, etc., making it difficult for frontline users such as university counselors and class advisors to quickly interpret key information; on the other hand, the systems lack personalized query functions (e.g., filtering students by student ID or major), result retrospective (e.g., viewing historical prediction records), and report export (e.g., standardized intervention recommendation reports), failing to achieve effective connection between "prediction results and intervention plans." While some universities have established mental health data management systems, their interfaces are complex and slow to respond, and they lack a dedicated module for "dynamic monitoring of depression trajectories," causing users to spend a significant amount of time processing data, significantly reducing the efficiency of mental health services.

[0008] In summary, existing technologies for predicting depression in college students have significant shortcomings in data integration, assessment paradigms, quality control, modeling methods, and interaction design, making it difficult to accurately predict the trajectory and potential changes in depression. Therefore, this invention proposes a prediction system for the trajectory and potential changes in depression in college students based on multi-source data. By integrating multimodal data, constructing a longitudinal tracking system, establishing a scientific data screening mechanism, employing multi-algorithm modeling and verification, and building a convenient interactive platform, this system aims to overcome the limitations of existing technologies and provide technical support for early intervention in college student depression. Summary of the Invention

[0009] In view of the technical problems mentioned in the background, an intelligent prediction system for the trajectory of depression changes in college students based on multimodal data fusion is provided.

[0010] The technical means employed in this invention are as follows:

[0011] A smart prediction system for the trajectory of depressive changes in college students based on multimodal data fusion includes: The system includes a multi-source data fusion and acquisition module, a label building unit, a model building and selection module, a real-time prediction unit, a result output unit, a web application module, and a privacy protection and compliance unit; among which, The multi-source data fusion acquisition module is used for initial data screening, multimodal collaborative calibration based on channel attention mechanism, and wavelet denoising processing to output clean data; The label construction unit is used to fit four-class long-term trajectory labels based on the potential category growth model and generate two-class short-term risk labels through potential transformation analysis. The model building and selection module is used to build an initial prediction model based on clean data and two types of labels, using five algorithms. After hyperparameter optimization through ten-fold cross-validation, the model with the best comprehensive performance is selected as the final prediction model using a preset index system. The real-time prediction unit is used to receive newly collected data, process it through standardization, and input it into the final prediction model to generate a depression change trajectory and risk prediction result with confidence score. The result output unit is used to visualize the prediction results and generate intervention suggestion templates, and to trigger early warning push for high-risk samples; The web application module is used to implement data uploading, result querying, and system configuration through a four-level hierarchical permission management system. The privacy protection and compliance unit is used to perform encrypted storage, encrypted transmission, and de-identification processing on data throughout the entire process, and to retain audit logs.

[0012] Furthermore, the multi-source data fusion acquisition module includes: an initial data acquisition unit, a longitudinal data acquisition unit, a data filtering unit, and a multimodal collaborative calibration subunit; The initial data acquisition unit collects basic information, demographic variables, depression-related questionnaire data, and resting-state EEG, near-infrared spectroscopy, MRI structural images, and MRI resting-state physiological data of freshmen. Based on the initial data collection, the longitudinal data collection unit organizes students to fill out the above three types of questionnaires every six months, covering eight time points from freshman to senior year. The multimodal collaborative calibration subunit employs a channel attention mechanism to adaptively learn the feature channel weights of questionnaire data and physiological data.

[0013] Furthermore, the label building unit verification process includes the following steps: Construct 1 sequentially Five latent category growth models, combined with Bayesian information criterion, entropy value ≥ 0.8, and Lo... Mendell Rubin likelihood ratio test P < 0.05 to select the optimal 4 subtypes; Bootstrap resampling 1000 times to verify stability, and labeling was completed when the sample proportion of each subtype and the difference coefficient of the growth curve parameters were all < 5%. Measurement invariance is verified by comparing the AIC and BIC values ​​of the fully measurement-invariant model and the partially measurement-invariant model. The transition probabilities of "low risk to medium / high risk" and "medium risk to high risk" are extracted, and short-term risk labels are generated with a threshold of ≥8%. The reliability of the labels is verified by AIC and BIC values ​​and classification accuracy of ≥85%. After the vertical update data is pushed, only short-term risk labels with a confidence level of <0.8 are dynamically corrected, and then synchronously updated to the long-term and short-term collaborative target label set.

[0014] Furthermore, the model construction and selection module also includes: Model selection: For binary classification labels, AUC is the primary indicator and weighted F1 score ≥ 85% is the secondary indicator; for four-class classification labels, weighted F1 score is the primary indicator and multi-class log loss < 0.5 is the secondary indicator; when multiple models meet the indicator thresholds, the accuracy on the independent validation set is compared, and the model with the highest accuracy is selected as the final prediction model. Model Iteration Unit: Every year, based on newly added longitudinal tracking data and intervention effect data, feature engineering and model training are re-executed, parameters are updated, and the model used in the real-time prediction unit is replaced.

[0015] Furthermore, the model building and selection module also includes a four-stage evaluation process: Phase 1: Construct an independent validation dataset, collect ≥1500 data points, preprocess them according to the feature engineering process, and label them with real labels; Phase 2: Evaluation metrics are defined as follows: for binary classification, there are 7 metrics including accuracy, precision, recall, F1 score, AUC, MSE, and MAE; for four-class classification, there are 10 metrics including accuracy, macro / weighted precision, macro / weighted recall, macro / micro / weighted F1 score, multi-class log loss, and confusion matrix accuracy. The third stage involves inputting the preprocessed independent validation set into the five-class model, outputting the prediction results and calculating the corresponding indicators. Phase 4: Conduct 10-fold cross-validation on the independent validation set, calculate the mean and standard deviation of the indicators, ensure that the standard deviation is less than 5%, and generate an evaluation report.

[0016] Furthermore, the Web application module comprises four sub-modules: The user login and registration submodule implements identity verification and four-level hierarchical access control. Students can view their own results, teachers can view class statistics, psychological counselors can view details of the students under their supervision and receive alerts, and administrators are responsible for system configuration and retain informed consent records through electronic forms. Data upload submodule: Supports uploading new student enrollment data and longitudinally collected data, automatically triggering quality control processes; Prediction Result Query Submodule: Supports filtering and querying by time range, name, student ID, and major, and displays visualized content and intervention suggestion templates; The system settings submodule supports administrator configuration of parameters, management permissions, model updates, and setting data collection reminder rules; it also addresses issues such as data collection failures and binary label prediction probabilities of 0.45. In abnormal scenarios such as 0.55, automatic supplementary sampling reminders and secondary prediction rules are configured. After the secondary prediction data is processed, it is input into the optimal model, and the output is taken with a confidence level ≥ 0.8. If all are < 0.8, manual review is triggered.

[0017] Compared with the prior art, the present invention has the following advantages: This invention improves data effectiveness to over 92% through multi-source data fusion and a three-level quality control mechanism. This avoids the bias issues associated with single data sources and leverages four years of longitudinal data collection to accurately capture dynamic changes, overcoming the limitations of traditional static assessments and providing comprehensive and reliable data support for mental health prediction. Simultaneously, the use of multi-algorithm modeling and a two-stage validation strategy ensures the optimal model achieves an external validation AUC ≥ 0.8 and maintains stable application across multiple universities, effectively overcoming the performance limitations and insufficient generalization capabilities of single algorithms and significantly improving prediction accuracy and scenario adaptability. Furthermore, the web-based visual interactive interface and hierarchical permission management design improve the efficiency of counselors obtaining assessment results by 60% and support data download and report export functions, achieving seamless integration of the "prediction-intervention" process. This greatly enhances the application efficiency and practical value of the technology, providing efficient and convenient technical support for early intervention in university mental health and helping universities build a scientific and normalized mental health management system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall system of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] like Figure 1 As shown, this invention provides an intelligent prediction system for the trajectory of depression changes in college students based on multimodal data fusion, including: a multi-source data fusion acquisition module, a label construction unit, a model construction and selection module, a real-time prediction unit, a result output unit, a web application module, and a privacy protection and compliance unit; The multi-source data fusion acquisition module is used for initial data screening, multimodal collaborative calibration based on channel attention mechanism, and wavelet denoising processing to output clean data; The label construction unit is used to fit four-class long-term trajectory labels based on the potential category growth model and generate two-class short-term risk labels through potential transformation analysis. The model building and selection module is used to build an initial prediction model based on clean data and two types of labels, using five algorithms. After hyperparameter optimization through ten-fold cross-validation, the model with the best comprehensive performance is selected as the final prediction model using a preset index system. The real-time prediction unit is used to receive newly collected data, process it through standardization, and input it into the final prediction model to generate a depression change trajectory and risk prediction result with confidence score. The result output unit is used to visualize the prediction results and generate intervention suggestion templates, and to trigger early warning push for high-risk samples; The web application module is used to implement data uploading, result querying, and system configuration through a four-level hierarchical permission management system. The privacy protection and compliance unit is used to perform encrypted storage, encrypted transmission, and de-identification processing on data throughout the entire process, and to retain audit logs.

[0023] Specifically, the system comprises a multi-source data fusion and acquisition module, a label building unit, a model building and selection module, a real-time prediction unit, a result output unit, a web application module, and a privacy protection and compliance unit, including the following processes: S1, the multi-source data fusion acquisition module performs data acquisition and optimization processing, and outputs clean data: S11. Integrating initial data acquisition, longitudinal data acquisition, data screening, and multimodal collaborative calibration subunits, invalid, contradictory, and missing samples are first removed through a three-level quality control mechanism to complete the initial data screening; S12. Based on the data after the initial screening in S11, the multimodal collaborative calibration subunit adopts a channel attention mechanism to adaptively learn the feature channel weights of the questionnaire data and physiological data, and assigns 60%-70% dynamic weights to physiological data such as EEG alpha band power spectral density and hippocampal cortex. S13. For the multimodal data calibrated by S12, wavelet denoising algorithm is used to process EEG signal noise and MRI head motion interference, complete data standardization and full-process quality control, and finally output clean data. S2. The tag construction unit uses clean data as input to build a set of long-term and short-term collaborative target tags: S21. Based on the cleaning data output by S1, extract the longitudinal BDI-II data from T0 to T7, and simultaneously retrieve the SCL-90 scale and college student mental health questionnaire data to supplement and improve the data. S22. Based on the dataset processed in S21, a quadratic growth curve is fitted using the latent class growth model (LCGM). The optimal number of categories is selected using the Bayesian information criterion (BIC) and entropy index to generate four-class long-term trajectory labels. S23. Combining the long-term trajectory labels generated in S22 with the original data in S21, the joint probability model is fitted using the expectation-maximization (EM) algorithm of potential transition analysis (LTA), and a 4×4 transition probability matrix of 7 adjacent time nodes is output to generate a binary short-term risk label. S24. The longitudinal data collection unit updates the questionnaire data every six months and pushes it to this unit to correct the short-term risk labels generated in S23, ultimately forming a set of long-term and short-term collaborative target labels. S3, the model building and selection module, uses clean data and two types of labels as input to build and select the optimal prediction model: S31. Using the clean data output from S1 and the two types of labels constructed in S2 as input, construct the initial prediction model based on five algorithms: logistic regression, random forest, k-nearest neighbors, support vector machine, and neural network. S32. In view of the characteristics of multimodal data, the neural network adopts the architecture of "feature-level cross-modal fusion + residual connection". First, the features of questionnaire text and physiological signal features are encoded separately, and then input into the residual network after being concatenated by the fusion layer. S33. For each initial model constructed in S31, the core hyperparameters are optimized using 10-fold cross-validation, with the iterative loss change being <10. -4 Stop training and save parameters. S34. For each model optimized in S33, a binary classification (AUC, F1 score, etc., 7 items) and a four-class classification (weighted F1 score, multi-class log loss, etc., 10 items) index system is used to quantitatively evaluate the model performance. S35. Based on the evaluation results of S34, select the model with the best overall performance as the final prediction model. Set up a model iteration unit. Every year, based on the newly added longitudinal tracking data and the intervention effect data reported by psychological counselors, conduct feature selection and model training again, and update the model parameters. S4. The real-time prediction unit processes new data and outputs prediction results: S41. Receive newly acquired data and transmit it to the multi-source data fusion acquisition module. Repeat the S11-S13 process to complete the standardization process and ensure that the data format is consistent with the model input requirements. S42. Input the new data processed in S41 into the optimal prediction model determined in S35 to generate the prediction results of the depression change trajectory and the risk of change. The prediction results are accompanied by a confidence score (≥0.8 is high confidence). S43. Simultaneously transmit the prediction results and confidence scores generated in S42 to the result output unit and the Web application module. S5. Result Output Unit, Web Application Module, and Privacy Protection and Compliance Unit perform result processing and control: S51. Receive the prediction results transmitted by S43. The result output unit visualizes the depression trajectory and risk level. For high-risk samples, generate an intervention suggestion template containing "risk level, intervention timing, intervention method, and follow-up period" and trigger an early warning push to the corresponding psychological counselor. The S52 and Web application modules connect the prediction results of S43 with the data collection link of S1. Through four-level hierarchical permission management, data uploading, result querying and system configuration are realized, and informed consent management functions are integrated synchronously. S53. For data throughout the entire process, the privacy protection and compliance unit adopts AES-256 encrypted storage, SSL transmission encryption and identity desensitization technology, and retains access audit logs for no less than one year. S54. Select an independent dataset with ≥1500 samples and verify the system performance according to the S1-S4 process. The results show that the weighted F1 value of the four-class label is ≥90% and the AUC of the two-class label is ≥0.88.

[0024] As a preferred embodiment, in this application, the initial data acquisition, longitudinal data acquisition, and data filtering units are specifically implemented as follows: The initial data collection unit collected basic information, demographic variables, depression-related questionnaire data, and physiological data from freshmen; the physiological data included resting-state EEG data, resting-state near-infrared spectroscopy data, MRI structural images, and MRI resting-state data. Based on the initial data collection, the longitudinal data collection unit organizes students to fill out the above three types of questionnaires every six months, covering eight time nodes from freshman to senior year (T0-T7), and the data is automatically synchronized to the data filtering unit. The data screening unit performs three levels of quality control purification: ① Verification of logical consistency of items, eliminating contradictory answer samples; ② Verification of completeness and rationality of answers, eliminating samples with a missing item rate >5% or answer time below the threshold (BDI-II ≥ 5 min, SCL-90 ≥ 10 min); ③ Cross-scale verification, eliminating samples with correlation coefficients below the threshold (BDI-II and SCL-90 depression factor r ≥ 0.6, College Student Mental Health Questionnaire "Depressive Tendency" and BDI-II r ≥ 0.5) and invalid manual review.

[0025] In this application, corresponding to S12-S13, the acquisition standards, quality control stratification, data management mechanism, and physiological data acquisition parameters of the multi-source data fusion acquisition module are as follows: The data collection standards set field formats and integrity verification rules for basic and demographic data; standardized scale versions, response time thresholds, and item logic verification rules for questionnaire data; and established equipment operation parameter standards and data collection process specifications for physiological data. The quality control process is implemented in three tiers: Level 1 is automated initial screening to remove samples with abnormal answer times, missing item rates >5%, or excessive physiological data artifacts; Level 2 is multi-dimensional verification to identify hidden anomalies; Level 3 is manual review to examine each suspected Level 2 sample. Data management establishes a three-tiered storage system: raw database, quality control database, and clean database. It records the entire process operation log and conducts regular quality audits. The raw data storage period does not exceed 3 years after the student's graduation, and the clean data is retained until 2 years after graduation. Upon expiration, the data is automatically de-identified and archived (the student ID is replaced with a random 6-digit code, and the middle character of the name is hidden). Physiological data acquisition parameters: Resting-state EEG was performed using a 32-channel device with a sampling rate of 250Hz, bandwidth of 0.1-100Hz, and electrode impedance ≤5kΩ; MRI data was performed using a 3.0T device with T1-weighted sequences TR=2500ms and TE=2.34ms, and data was re-acquired when head movement displacement was >2mm.

[0026] As a preferred option, corresponding to S22-S24, the specific parameters and verification process for tag construction are as follows: LCGM models of categories 1-5 were constructed sequentially. The optimal 4 subtypes were selected by combining BIC, entropy (≥0.8), and Lo-Mendell-Rubin likelihood ratio test (LMR, P<0.05). Bootstrap resampling (sampling times=1000) was used to verify stability. Labeling was completed when the sample proportion of each subtype and the difference coefficient of the growth curve parameters were all <5%. Measurement invariance is verified by comparing the AIC and BIC values ​​of the fully measurement invariant model and the partially measurement invariant model. The transition probabilities of "low risk to medium / high risk" and "medium risk to high risk" are extracted, and short-term labels are generated with a threshold of ≥8%. The reliability of the labels is verified by AIC, BIC values ​​and classification accuracy (≥85%). After the vertical update data is pushed, only short-term risk labels with a confidence level of <0.8 are dynamically corrected, and then synchronously updated to the long-term and short-term collaborative target label set.

[0027] As a preferred embodiment, in this application, corresponding to S31-S33, the range of core hyperparameters for predictor variable preprocessing, feature engineering, and the five algorithms is as follows: Predictor variable preprocessing and feature engineering steps: ① Determine the four categories of original predictor variables: basic information, demographics, questionnaires, and physiology; ② Remove outlier samples with a missing rate > 5% and those judged by the 3σ criterion; fill variables with a missing rate ≤ 5% using the mean (numerical) or mode (categorical); ③ One-hot encoding for unordered categorical variables, ordinal encoding for ordered categorical variables, and Z-score standardization for numerical variables; ④ Use "10-fold cross-validation recursive feature elimination (RFE) + random forest feature importance ranking" to select core features, retain variables with a contribution ≥ 10%, and select the top 80% of features to form the final feature set; The dataset was divided into a cross-validation set and an independent test set using 8:2 stratified sampling. The cross-validation set was further divided into 10 subsets for 10-fold cross-validation. The optimization range of the core hyperparameters for the five algorithms: Logistic Regression regularization coefficient λ∈{0.01,0.1,1,10}, learning rate∈{0.001,0.01,0.1}; Random Forest tree count∈{50,100,200}, maximum depth∈{5,10,15,None}; k-nearest neighbor algorithm k value∈{5,10,15,20}; Support Vector Machine kernel function parameter γ∈{0.01,0.1,1,10}, penalty coefficient C∈{1,5,10,20}; Neural Network hidden layer first layer neuron count∈{32,64,128}, Dropout probability∈{0.1,0.2,0.3}, learning rate∈{0.0001,0.001,0.01}.

[0028] In a preferred embodiment, in this application, corresponding to S35, the specific execution of the model selection and iteration unit is as follows: For binary classification labels, AUC is the primary indicator and weighted F1 score (≥85%) is the secondary indicator; for four-class classification labels, weighted F1 score is the primary indicator and multi-class log loss (<0.5) is the secondary indicator. When multiple models meet the indicator threshold, compare the accuracy of the independent validation set and select the one with the highest accuracy as the final prediction model. Every year, the model iteration unit re-executes feature engineering and model training based on newly added longitudinal tracking data and intervention effect data, updates parameters, and replaces the model used in the real-time prediction unit.

[0029] In a preferred embodiment, in this application, corresponding to S34, the four-stage evaluation process of the model verification unit is as follows: The first stage involves constructing an independent validation dataset, collecting ≥1500 data points, preprocessing them according to the S3 feature engineering workflow, and labeling them with real labels based on the LCGM and LTA methods. The second phase of evaluation metrics includes: seven items for binary label fitting accuracy, precision, recall, F1 score, AUC, MSE, and MAE; and ten items for four-category label fitting accuracy, macro / weighted precision, macro / weighted recall, macro / micro / weighted F1 score, multi-category log loss, and confusion matrix accuracy. The third stage involves inputting the preprocessed independent validation set into the five-class model, outputting the prediction results and calculating the corresponding indicators. In the fourth stage, ten-fold cross-validation is carried out on the independent validation set, the mean and standard deviation of the indicators are calculated, the standard deviation is ensured to be <5%, and an evaluation report is generated.

[0030] In a preferred embodiment, in this application, corresponding to S52, the specific functions of the four sub-modules of the Web application module are as follows: The user login and registration submodule implements identity verification and four-level hierarchical access control: students can view their own prediction results and suggestions, but are prohibited from downloading and forwarding; teachers can view the group statistics of their class, but are prohibited from viewing individual privacy; psychological counselors can view the details of the students under their supervision and receive warnings, but are prohibited from exporting in batches; administrators are responsible for system configuration and are prohibited from accessing user data and prediction results; student consent is obtained through electronic forms, and informed consent records are retained for future reference. The data upload submodule supports uploading data of newly enrolled students and longitudinally collected data, which is automatically synchronized to the multi-source data fusion acquisition module and triggers the S11 quality control process. The prediction results query submodule supports filtering by time range, name, student ID, and major, and displays S51 visualization content and intervention suggestion templates; The system settings submodule supports administrators in configuring parameters, managing permissions, updating models, and setting collection reminder rules. For abnormal scenarios such as data collection failure and binary label prediction probability of 0.45-0.55, it configures automatic supplementary collection reminders and secondary prediction rules. The secondary prediction data is processed by S41 and then input into the optimal model. The output is the result with a confidence level ≥0.8. If the confidence level is <0.8, manual review is triggered.

[0031] Example 1 Using a comprehensive university (4 faculties, with an average of 3,000 new students per year) as the research subject, starting in September 202X, we have continuously carried out four years of longitudinal data accumulation, model construction and verification, and practical application to achieve accurate prediction of the depressive trajectory of freshmen enrolled in 202X+1 and subsequent years, providing data support and decision-making basis for graded intervention of mental health in universities.

[0032] As an embodiment of this application, the specific implementation steps are as follows: 1. Implementation of multi-source data fusion acquisition module Initial data collection unit (T0: within 1 week of new students' enrollment) Basic / demographic data: Age (16-22 years old), major (matching the "Major Catalog" code) and other data were collected through the campus welcome system. Field validation was set (non-integer age pop-up correction). 2986 valid data entries were obtained (validity rate 99.5%).

[0033] Questionnaire data: The encrypted platform distributed BDI-II (21 items) and SCL-90 (domestic revised version), etc., and controlled the answering time (BDI-II ≥ 5 min, SCL-90 ≥ 10 min). Simultaneously preset the logical consistency verification rules of the questions, and collected 2978 questionnaires. 12 abnormal samples were removed (effective rate 99.7%).

[0034] Physiological data: Standardized data collection was completed in a university mental health laboratory. The equipment and operating parameters strictly followed the preset standards: ① Resting-state EEG data: NeuroScan32 EEG equipment was used, with a sampling rate of 250Hz, bandwidth of 0.1-100Hz, and electrode impedance ≤5kΩ, resulting in 2950 effective data collections; ② Resting-state near-infrared data: An 8-channel device was used, with the data artifact ratio controlled to ≤20%, resulting in 2914 effective data collections; ③ MRI data: A 3.0T MRI device was used, with T1-weighted sequences TR=2500ms and TE=2.34ms. Data was re-collected when head movement displacement was >2mm, resulting in 2896 effective data collections.

[0035] Longitudinal data collection units (T1-T7: once every six months): After the collection was completed at T0, longitudinal tracking collection was carried out once every six months from freshman to senior year, completing data collection at seven time nodes: T1 (freshman year, second semester), T2 (sophomore year, first semester), T3 (sophomore year, second semester), T4 (junior year, first semester), T5 (junior year, second semester), T6 (senior year, first semester), and T7 (senior year, second semester). The questionnaire collection process and quality control rules of T0 were reused at each node. Follow-up compliance was ensured through multiple channels such as campus notifications, counselor reminders, and push of exclusive questionnaire links. Finally, 2,820 valid follow-up results were completed at node T7, with a follow-up rate of 94.4% throughout the process.

[0036] 2. Implementation of the data quality control module Level 1 automated initial screening: Python scripts removed samples with a questionnaire missing rate >5%, excessive physiological data artifacts, and response time below the threshold, totaling 242 samples.

[0037] Secondary multidimensional verification: Cross-scale validation and data correlation verification were performed. The correlation coefficient of the questionnaire (BDI-II and SCL-90 depression factor r=0.68) and the correlation of physiological indicators (prefrontal cortex thickness and α band r=-0.42) were calculated. 23 samples were included in the "suspicion database" and 26 samples were marked for verification.

[0038] Three-level manual review: 3 associate professors of psychology + 2 analysts reviewed the data, 15 perfunctory answers were removed, 7 abnormal equipment samples were re-collected, and finally 2786 clean data were obtained (effectiveness rate 93.2%).

[0039] Data management is implemented across the entire process: a standardized CSV storage format and a "time-type-number" naming rule are adopted to build a three-level storage system of raw database, quality control database, and cleaning database. The system fully records the entire process of data collection, screening, cleaning, and verification. A data quality audit is conducted once a quarter to ensure that the data is traceable.

[0040] 3. Implementation of the Tag and Model Building Module 1. Label Construction: Modeling was performed using Mplus 8.6. The LCGM was used to divide the depression trajectory into four categories (stable low risk 68.3%, slowly rising 21.4%, fluctuating and recurring 6.4%, and rapidly deteriorating 4.0%). After Bootstrap verification to ensure stability, four-category labels were generated. The LTA algorithm output a transition probability matrix of 7 nodes (core: stable low risk → slowly rising 12.3%). A binary transition risk label was generated with an 8% threshold (325 items with risk and 2461 items without risk, with a classification accuracy of 86.2%).

[0041] 2. Model Construction: After cleaning, encoding standardization and screening using "RFE + Random Forest" on 124 original features, 86 core features were determined; 2786 clean data points were stratified into a cross-validation set (2229 data points) and an independent test set (557 data points) in an 8:2 ratio, and training and optimization of five types of models (LR, RF, KNN, SVM, and NN) were completed through 10-fold cross-validation.

[0042] 3. Model Validation and Selection: Two-stage validation showed that the RF model was optimal (binary classification AUC 0.895, accuracy 89.6%; four-class weighted F1 score 87.8%, log loss 0.45), and the standard deviation of the 10-fold cross-validation index on the independent validation set was <5%. Therefore, RF was finally selected as the optimal model.

[0043] 4. Implementation of the real-time prediction and results application module Real-time prediction: 2,960 data points were collected from the freshmen of the Class of 202X+1. After cleaning, 2,892 data points were input into the RF model, which automatically outputs the four-category depression trajectory prediction results and the two-category conversion risk prediction results for each freshman. The accuracy rate was 88.6% after 10% sample verification.

[0044] Web application: Front-end and back-end separation architecture (Vue.js + SpringBoot), implementing hierarchical permission management, data upload (1000 records per time), visualization (line chart / heatmap), report download, supporting 500 concurrent users, with a response time of <1.5 seconds.

[0045] Intervention results: 498 high-risk students received tiered intervention. After 6 months, 68.1% of the "rapid deterioration type" improved to mild, the mean BDI-II decreased from 16.2 to 11.5, and the efficiency of counselor inquiries increased by 80%.

[0046] In summary, this embodiment verifies the feasibility of the system: the data effectiveness rate is 93.2%, the external validation accuracy rate of the model is 87.2%, the improvement rate of high-risk students is 68.1%, it overcomes the limitations of existing technologies, the technical parameters can be adjusted as needed, and it is suitable for different university scenarios.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent prediction system for the trajectory of depressive changes in college students based on multimodal data fusion, characterized in that, include: The system includes a multi-source data fusion and acquisition module, a label building unit, a model building and selection module, a real-time prediction unit, a result output unit, a web application module, and a privacy protection and compliance unit. The multi-source data fusion acquisition module is used for initial data screening, multimodal collaborative calibration based on channel attention mechanism, and wavelet denoising processing to output clean data; The label construction unit is used to fit four-class long-term trajectory labels based on the potential category growth model and generate two-class short-term risk labels through potential transformation analysis. The model building and selection module is used to build an initial prediction model based on clean data and two types of labels, using five algorithms. After hyperparameter optimization through ten-fold cross-validation, the model with the best comprehensive performance is selected as the final prediction model using a preset index system. The real-time prediction unit is used to receive newly collected data, process it through standardization, and input it into the final prediction model to generate a depression change trajectory and risk prediction result with confidence score. The result output unit is used to visualize the prediction results and generate intervention suggestion templates, and to trigger early warning push for high-risk samples; The web application module is used to implement data uploading, result querying, and system configuration through a four-level hierarchical permission management system. The privacy protection and compliance unit is used to perform encrypted storage, encrypted transmission, and de-identification processing on data throughout the entire process, and to retain audit logs.

2. The intelligent prediction system for the trajectory of depressive changes in college students based on multimodal data fusion according to claim 1, characterized in that, The multi-source data fusion acquisition module includes: an initial data acquisition unit, a longitudinal data acquisition unit, a data filtering unit, and a multimodal collaborative calibration subunit; The initial data acquisition unit collects basic information, demographic variables, depression-related questionnaire data, and resting-state EEG, near-infrared spectroscopy, MRI structural images, and MRI resting-state physiological data of freshmen. Based on the initial data collection, the longitudinal data collection unit organizes students to fill out the above three types of questionnaires every six months, covering eight time points from freshman to senior year. The multimodal collaborative calibration subunit employs a channel attention mechanism to adaptively learn the feature channel weights of questionnaire data and physiological data.

3. The intelligent prediction system for the trajectory of depressive changes in college students based on multimodal data fusion according to claim 1, characterized in that, The tag building unit verification process includes the following steps: Construct 1 sequentially Five latent category growth models, combined with Bayesian information criterion, entropy value ≥ 0.8, and Lo... Mendell Rubin likelihood ratio test P < 0.05 to select the optimal 4 subtypes; Bootstrap resampling 1000 times to verify stability, and labeling was completed when the sample proportion of each subtype and the difference coefficient of the growth curve parameters were all < 5%. Measurement invariance is verified by comparing the AIC and BIC values ​​of the fully measurement invariant model and the partially measurement invariant model. The transition probabilities of "low risk to medium / high risk" and "medium risk to high risk" are extracted, and short-term risk labels are generated with a threshold of ≥8%. The reliability of the labels is verified by AIC and BIC values ​​and classification accuracy of ≥85%. After the vertical update data is pushed, only short-term risk labels with a confidence level of <0.8 are dynamically corrected, and then synchronously updated to the long-term and short-term collaborative target label set.

4. The intelligent prediction system for the trajectory of depressive changes in college students based on multimodal data fusion according to claim 1, characterized in that, The model building and selection module also includes: Model selection: For binary classification labels, AUC is the primary indicator and weighted F1 score ≥ 85% is the secondary indicator; for four-class classification labels, weighted F1 score is the primary indicator and multi-class log loss < 0.5 is the secondary indicator; when multiple models meet the indicator thresholds, the accuracy on the independent validation set is compared, and the model with the highest accuracy is selected as the final prediction model. Model Iteration Unit: Every year, based on newly added longitudinal tracking data and intervention effect data, feature engineering and model training are re-executed, parameters are updated, and the model used in the real-time prediction unit is replaced.

5. The intelligent prediction system for the trajectory of depressive changes in college students based on multimodal data fusion according to claim 1, characterized in that, The model building and selection module also includes a four-stage evaluation process: Phase 1: Construct an independent validation dataset, collect ≥1500 data points, preprocess them according to the feature engineering process, and label them with real labels; Phase 2: Evaluation metrics are defined as follows: for binary classification, there are 7 metrics including accuracy, precision, recall, F1 score, AUC, MSE, and MAE; for four-class classification, there are 10 metrics including accuracy, macro / weighted precision, macro / weighted recall, macro / micro / weighted F1 score, multi-class log loss, and confusion matrix accuracy. The third stage involves inputting the preprocessed independent validation set into the five-class model, outputting the prediction results and calculating the corresponding indicators. Phase 4: Conduct 10-fold cross-validation on the independent validation set, calculate the mean and standard deviation of the indicators, ensure that the standard deviation is less than 5%, and generate an evaluation report.

6. The intelligent prediction system for the trajectory of depressive changes in college students based on multimodal data fusion according to claim 1, characterized in that, The web application module comprises four sub-modules: The user login and registration submodule implements identity verification and four-level hierarchical access control. Students can view their own results, teachers can view class statistics, psychological counselors can view details of the students under their supervision and receive alerts, and administrators are responsible for system configuration and retain informed consent records through electronic forms. Data upload submodule: Supports uploading new student enrollment data and longitudinally collected data, automatically triggering quality control processes; Prediction Result Query Submodule: Supports filtering and querying by time range, name, student ID, and major, and displays visualized content and intervention suggestion templates; The system settings submodule supports administrator configuration of parameters, management permissions, model updates, and setting data collection reminder rules; it also addresses issues such as data collection failures and binary label prediction probabilities of 0.

45. In abnormal scenarios such as 0.55, automatic supplementary sampling reminders and secondary prediction rules are configured. After the secondary prediction data is processed, it is input into the optimal model, and the output is taken with a confidence level ≥ 0.

8. If all are < 0.8, manual review is triggered.

Images