Teenager psychological sub-health intelligent early warning system based on multi-source heterogeneous data fusion

Abstract

The invention discloses a teenager psychological sub-health intelligent early warning system based on multi-source heterogeneous data fusion, and the system comprises a data collection layer which collects the behavior, physiological and social three-dimensional data of teenagers through the multi-source channels of campus cards, wearable devices, social media and questionnaires, and builds an original data pool through cleaning, denoising and standardization; the data fusion layer is used for integrating multi-source data by adopting weighted average and Kalman filtering, mining psychological sub-health key indexes in combination with a feature selection and extraction technology, and forming a three-dimensional psychological portrait; the deep learning layer is used for constructing a multi-modal fusion early warning model based on a Transform architecture, and carrying out real-time prediction and dynamic tracking of psychological sub-health risks through large-scale data training and cross validation optimization; and the intelligent early warning layer is used for visually displaying an early warning result, integrating three-party linkage of a management end, a teacher end and a parent end, providing 24-hour online intervention by a built-in AI psychological counseling module, and automatically transferring high-risk cases to professional psychological consultants.

Description

Technical Field

[0001] This invention relates to the field of crisis early warning and psychological services for adolescents, and in particular to an intelligent early warning system for adolescent sub-health based on the fusion of multi-source heterogeneous data. Background Technology

[0002] The current serious social situation regarding the mental health of adolescents has drawn widespread attention. The World Health Organization points out that 75% of adult mental illnesses begin before the age of 24, and if effective intervention is not implemented in the early stages, it will lead to chronic mental illness and a continuous decline in quality of life. Therefore, building a three-tiered support network of "family-school-community," promoting social-emotional education courses, and strengthening professional psychological intervention have become urgent priorities—this is not only to protect individual growth but also a fundamental investment in social stability and long-term development.

[0003] However, traditional mental health assessment methods are static, have delayed interventions, and lack sufficient integration of multi-source data, making it difficult to accurately identify mental health risks in adolescents. Traditional assessments rely on static data from questionnaires and interviews, which cannot dynamically track changes in mental state. Furthermore, they are susceptible to data distortion due to social desirability bias and test fatigue. Even when risks are identified, subsequent interventions are often delayed due to information lag and fragmented resources, resulting in approximately 76% of mental health problems worsening due to lack of timely intervention. For example, Xiaoming, a left-behind adolescent, exhibits sub-health symptoms such as depression and social withdrawal due to a long-term lack of parent-child interaction. In traditional assessments, Xiaoming might mask his true emotions in the questionnaire due to social desirability, or his fluctuating state during the assessment might lead to biased results, making it impossible to accurately identify the risks. Therefore, this paper proposes an intelligent early warning system for adolescent mental health sub-health based on the fusion of multi-source heterogeneous data. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent early warning system for adolescent mental sub-health based on multi-source heterogeneous data fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart early warning system for adolescent mental health sub-health based on multi-source heterogeneous data fusion includes: Data Acquisition Layer: Collects three-dimensional data on adolescent behavior, physiology, and social behavior through multiple heterogeneous channels, including campus card consumption records, physiological data monitored by wearable devices, social media language analysis, and questionnaire surveys. After data cleaning, noise reduction, and standardization, a raw data pool is constructed. Data fusion processing layer: The weighted average and Kalman filter fusion algorithms are used to integrate multi-source data. Combined with feature selection and extraction techniques, key predictive indicators of sub-psychological health are mined from the fused high-dimensional feature space to form a three-dimensional psychological profile to support model analysis. Deep learning early warning model layer: Based on the Transformer deep learning architecture, a multimodal fusion early warning model is constructed. The model is input through data preprocessing and feature engineering, trained on a large-scale labeled dataset, and the parameters are optimized through cross-validation. Ensemble learning is then used to perform real-time prediction and dynamic tracking of psychological sub-health risks. Intelligent early warning and intervention platform layer: The early warning results are displayed visually on mobile devices. It integrates a three-way linkage mechanism between the management end, the teacher end, and the parent end. The built-in AI psychological counseling module provides 24-hour online intervention, and high-risk cases are automatically referred to professional psychological counselors.

[0006] The above technical solution further includes: Furthermore, the construction of the original data pool includes the following steps: By connecting to the school's campus card system API, structured data such as campus card consumption records (time / location / amount), access control system entry and exit times, and library borrowing frequency are obtained. With student authorization, time-series data such as heart rate, sleep quality, and steps are synchronized through wearable device manufacturer SDKs (such as nighttime heart rate variability collected by smart bracelets). Under the premise of compliance, Python web scraping frameworks (such as Scrapy) are used to collect publicly available social media text, emojis, and interactive behaviors. After filtering out noise such as advertisements using regular expressions, the data is stored in JSON format. Traditional scales such as SCL-90 and MHT are transformed into online questionnaires. Self-reported results of depression and anxiety are collected electronically through a psychological assessment system and linked to student IDs to form a basic psychological profile. The depression risk level (low / medium / high) is labeled based on the results of psychological assessments, and the sentiment tendency (positive / neutral / negative) of social media texts is labeled using NLP models to form a structured labeling system; Tiered storage: Student Basic Information Table: Stores static information such as student ID, gender, age, grade, left-behind status, and frequency of parent-child communication, supporting subsequent group analysis and risk factor association research; Multi-source data master table: behavioral, physiological and social data are associated by timestamp, including raw data, cleaned data and annotation results, forming the core data base for dynamic tracking; Risk warning label table: Records warning signals (such as depressive tendencies), warning levels and intervention status output by the model, supporting real-time warnings and closed-loop interventions.

[0007] Furthermore, the integration of multi-source data using a weighted average and Kalman filter fusion algorithm includes the following steps: The weighted average weighting is based on prior research or expert knowledge to assess the degree of influence of each feature on the psychological state. For example, negative sentiment on social media may be given a higher weight (e.g., 0.3), while fluctuations in consumption frequency may be given a lower weight (e.g., 0.1). The sum of the weights of all features is 1. Kalman filtering dynamic fusion is adopted to define a state vector containing psychological state indicators such as depression level and anxiety level, such as [depression score, anxiety score]. Based on the pattern of psychological state change, a state transition equation is established to describe the trend of state vector change over time. Multi-source features are mapped to the state vector, such as associating social media sentiment score with depression score. Initialize the state vector and covariance matrix to reflect the estimation uncertainty at the initial time step. Predict the current state vector and covariance matrix using the state transition matrix. Calculate the Kalman gain and adjust the predicted state by combining it with weighted average observations (such as weighted social media sentiment scores). Update the state estimate and covariance matrix to provide a basis for the prediction of the next time step, forming a dynamic closed loop.

[0008] It outputs accurate estimates of psychological state indicators such as depression level and anxiety level at the current moment, forming a dynamic psychological profile. It quantifies the uncertainty of the estimate through covariance matrix, reflects the correlation and confidence interval between various state indicators, and triggers psychological sub-health warnings (such as generating a high-risk signal when the depression score is >0.7) based on the comparison between the state estimation results and preset thresholds, providing a basis for subsequent intervention.

[0009] Furthermore, the method of combining feature selection and extraction techniques to mine key predictive indicators of sub-optimal mental health from the fused high-dimensional feature space includes the following steps: The data from four dimensions—campus card spending records, wearable device physiological data, social media texts, and psychological assessment results—are integrated into a high-dimensional feature matrix that includes numerical (spending frequency), time-series (heart rate fluctuations), textual (emotional tendencies), and subjective (depression scores). Using individual adolescents as rows and feature dimensions as columns, a matrix is ​​constructed that includes features such as consumption patterns, changes in physiological indicators, and emotional tendencies on social media. A chi-square test is performed on each feature and its correlation with sub-health states (such as depression / anxiety) to calculate the p-value to assess the correlation between the feature and the state. For example, the frequency of use of negative emotion words on social media may show a significant correlation with depression (p<0.05), while fluctuations in consumption frequency may have a weaker correlation. A p-value threshold (such as 0.05) is set to select a subset of features that are significantly related to sub-health states as input for subsequent feature extraction. Linear transformation is performed on the filtered feature matrix to extract the top N principal components (such as principal components with a cumulative variance contribution rate > 85%), which preserves the overall data structure while reducing dimensionality. For example, principal component analysis (PCA) is used to integrate features such as consumption frequency, heart rate variability, and social media sentiment into three principal components: “behavioral stability”, “physiological state fluctuation”, and “social sentiment”. For the task of distinguishing between sub-healthy and non-sub-healthy mental states, linear discriminant analysis (LDA) is used to find the discriminative direction, maximize the inter-class variance, minimize the intra-class variance, and extract discriminative feature vectors. Features extracted from PCA, LDA, and self-attention mechanisms are integrated into three dimensions: behavior, physiology, and social. The behavior dimension includes participation in campus activities and regularity of daily routines; the physiology dimension includes heart rate variability and sleep quality; and the social dimension includes frequency of social media interactions and emotional tendencies. Based on the integrated features, a three-dimensional psychological profile containing specific values ​​or categories is generated for each adolescent. For example, the profile may show that a student has a "behavioral dimension score of 0.8 (high activity level), physiological dimension score of 0.6 (normal fluctuation), and social dimension score of 0.3 (negative tendency)", and the overall assessment is "low risk". Through heatmaps, radar charts, or dashboards, the behavioral, physiological, and social characteristics and overall risk levels can be presented intuitively. For example, radar charts can show the distribution of students' scores in each dimension, while heatmaps can show the characteristic correlation patterns at the group level.

[0010] Furthermore, the construction of the multimodal fusion early warning model based on the Transformer deep learning architecture includes the following steps: The text modality uses a pre-trained BERT model to encode social media text, generating sentiment vectors (e.g., positive / neutral / negative) to capture emotional features in social interactions. The physiological modality uses a sliding window to extract statistical features (e.g., 7-day average heart rate, deep sleep ratio) from time-series data such as heart rate and sleep quality. The behavioral modality performs frequency statistics and amount distribution analysis on campus card spending records to generate consumption pattern feature vectors (e.g., high-frequency small-amount spending tendency). The subjective modality converts psychological assessment scores into standardized values ​​to form continuous variables of depression / anxiety tendency. An independent Transformer encoder is designed for each modality (behavioral / physiological / social / subjective). Each layer contains a multi-head self-attention mechanism and a feedforward neural network to capture local and global features within the modality. For example, the social text encoder focuses on "negative emotion words" through self-attention, while the physiological data encoder captures heart rate variability trends. A cross-modal attention mechanism is used at the encoder output to dynamically weight and fuse feature vectors from different modalities. For example, the "negative emotion" feature of social text and the "heart rate variability" feature of physiological data are associated through attention weights to form a fused feature vector. We use transfer learning to initialize model weights, combine data augmentation techniques (such as text synonym replacement and physiological data noise injection) to improve generalization ability, and optimize hyperparameters such as learning rate, batch size, and number of attention heads through grid search to achieve optimal performance on the validation set. The model output includes the probability of sub-optimal mental health risk, risk level (low / medium / high), and the ranking of the importance of key features (such as 30% for negative tendencies in social texts and 25% for abnormal physiological data), supporting intervention decision-making.

[0011] Furthermore, the real-time prediction and dynamic tracking of psychological sub-health risks includes the following steps: The preprocessed feature vectors are captured by a self-attention layer to capture cross-modal associations (such as the association between social text "negative emotions" and physiological data "heart rate variability"), and the risk probability is output through a fully connected layer (such as 0.8 indicating high risk). The model output includes the risk level (low / medium / high), probability value and contribution of key features (such as social modality accounting for 40%), and is transmitted to the early warning platform. The risk trends of individuals / groups are dynamically presented through heat maps and line graphs. For example, if a student's risk probability increases from 0.3 to 0.7 in the past 7 days, high-risk cases will automatically trigger APP push, SMS and email, and be synchronized to the management terminal, teacher terminal and parent terminal, start the "AI psychological counseling module" for 24-hour emotional support, and refer to professional psychological counselors. By continuously collecting behavioral, physiological, and social data through wearable devices and campus systems, an individual psychological state time series database is constructed, recording historical changes in state (such as fluctuations in sleep quality from the beginning to the end of the semester), generating a psychological profile report every month, and forming a dynamic risk profile by combining SCL-90 scale scores, model prediction results, and intervention effects. Kalman filtering is used to predict psychological state trends. Abnormal fluctuations (such as a sudden 20% drop in heart rate variability) are detected by the 3σ principle, triggering a primary warning. Thresholds are set according to risk probability (such as >0.6 triggering a medium warning and >0.8 triggering a high warning), and the intensity of intervention is automatically adjusted (such as increasing the frequency of AI tutoring and referring to professional counseling).

[0012] Furthermore, the aforementioned tripartite linkage mechanism between the management end, the teacher end, and the parent end includes: Standardized interfaces are built using RESTful API / GraphQL protocols to support the access of heterogeneous data, such as campus behavioral data (e.g., consumption records), physiological data (e.g., heart rate), and social media text (analyzed by NLP sentiment analysis), ensuring HTTPS encrypted transmission and compliance with the "Regulations on the Protection of Children's Personal Information Online". Sensitive data (such as student names) is anonymized. The management side can access the school-wide aggregated data and individual risk profiles, while the teachers side can only access the behavior and early warning information of the students in their classes, and the parents side can only access their own children's physiological and emotional reports, forming a three-level data view of macro, meso and micro. Integrate third-party messaging services and customize push content according to roles—the management side receives the school's high-risk list and intervention effect report, the teacher side receives warnings of students' emotional fluctuations and suggested measures (such as after-school counseling), and the parent side receives reminders of their children's abnormal emotions and family care guidelines (such as sleep improvement suggestions). The built-in video call function supports real-time communication between teachers and parents. Voice-to-text conversion automatically generates communication records and archives them in the student's psychological profile, ensuring that the communication content is traceable. Based on RBAC, the access control system defines three roles: administrator (system configuration / data export), teacher (intervention work order creation / class data viewing), and parent (family information update / children's report viewing). Permission changes are controlled through OA approval flow to ensure compliance. All roles' access to and operations on sensitive data (such as modification of warning thresholds) are recorded. Logs are stored on the blockchain to ensure immutability and support filtering and backtracking by time, user, and operation type. High-risk students automatically generate work orders and assign responsibility to relevant parties—for mild risks, a notification is sent to the teacher's end suggesting classroom attention; for moderate risks, a notification is sent to the management end to coordinate the intervention of a psychological counselor; for severe risks, an emergency process is triggered to notify parents and coordinate with the school doctor for on-site assessment. The work order status is updated in real time, supporting the marking of processing progress and results by all three parties. The effects of different measures are compared through A / B testing (such as teacher counseling vs. family care). Evaluation indicators include the rate of improvement in mood, changes in academic performance, and the number of re-warnings.

[0013] Furthermore, the built-in AI psychological counseling module provides 24-hour online intervention, including the following steps: By integrating voice, text, physiological signals (such as heart rate / skin conductance activity from wearable devices), and behavioral data (such as campus card spending patterns), NLP is used to extract semantic features from the text (such as BERT model to analyze emotion labels), CNN is used to process facial expression images, and LSTM is used to analyze speech spectral features (such as fundamental frequency and energy) to form a structured feature vector. The features of each modality are dynamically weighted through the Transformer self-attention mechanism to construct a comprehensive representation vector of the user's psychological state, supporting subsequent emotion computing and risk assessment. Based on a pre-trained BERT-NLP model, emotion labels (such as "anxiety -0.85" and "sadness -0.72") are identified. Combined with a keyword library (such as "suicide" and "despair"), emergency alerts are triggered. A bidirectional LSTM classifier is used to analyze features such as trembling tone and abnormal speech rate to help judge the intensity of emotions and stress status. Heart rate variability (HRV) is monitored in real time through wearable devices. When HRV decreases, stress status is indicated. Combined with behavioral data (such as nighttime consumption frequency), the risk is comprehensively assessed. The fused features are input into a deep neural network (DNN) to output the depression / anxiety risk level (low / medium / high) and intervention priority. The assessment results are dynamically updated every 15 minutes. When the risk level rises to "medium / high", the system automatically pushes an intervention pop-up to the user (such as "You look a little tired lately, would you like to talk to me?"), guiding the user into an encrypted chat interface. The user actively inputs text / voice through the "emotional tree hole" function, and the system analyzes the emotions in real time and generates a response, supporting multi-round dialogue to adjust the strategy; Layered response: For low-risk users: Push cognitive behavioral therapy (CBT) self-help tools (such as emotion diary templates) and mindfulness meditation audio to guide them in self-regulating their emotions; For medium-risk users: Initiate a structured dialogue, guide users to describe their specific concerns, provide a solution library (such as "How to deal with test anxiety?"), and generate personalized suggestions by combining knowledge graphs (such as the DSM-5 diagnostic criteria); High-risk users: Generate emergency intervention scripts (such as "You are in danger right now, I need to contact your teacher / parents, is that okay?"), and simultaneously trigger the manual referral process; It uses the Tacotron 2 speech synthesis model to generate natural speech responses, supports dialect adaptation, and uses 3D face reconstruction technology to generate virtual counselor expressions, matching the response emotions (such as a smile to comfort, a serious reminder) to enhance the interactive experience; If a user explicitly expresses suicidal intent (e.g., "I want to jump off a building"), and the emotion computing model determines "extreme anxiety" five times in a row, an emergency referral is triggered. If the user refuses all self-help intervention suggestions and the conversation lasts for more than 30 minutes without alleviating the emotion, the intervention is upgraded to manual intervention. The system pushes pop-ups to the management and teachers' terminals, displaying the user's risk level, last interaction record, and suggested referral recipients (such as school counselors). Counselors can view user data through an encrypted channel within 5 minutes and decide whether to take over. During referral, only necessary information (such as risk level and key dialogue excerpts) is shared. User identity is anonymized, in accordance with the "Regulations on the Protection of Children's Personal Information Online". AI automatically generates referral case reports, including emotional change curves and the effectiveness of intervention measures. Counselors evaluate the referral results and optimize the risk assessment model thresholds.

[0014] The present invention has the following beneficial effects: This invention employs multi-source heterogeneous channels to collect behavioral, physiological, and social three-dimensional data. After cleaning and denoising, an original data pool is constructed. Weighted averaging and Kalman filtering algorithms are used to integrate the multi-source data, and feature selection is combined to extract key indicators of sub-psychological health, forming a three-dimensional psychological profile to support dynamic tracking. A multimodal fusion model is constructed based on the Transformer architecture, trained and integrated with large-scale labeled data, to perform real-time prediction and dynamic tracking of sub-psychological health risks, overcoming data distortion caused by social desirability bias and test fatigue. The intelligent early warning and intervention platform layer displays early warning results visually on mobile devices, integrates a three-way linkage mechanism between management, teachers, and parents, and has a built-in AI psychological counseling module to provide 24-hour online intervention. High-risk cases are automatically referred to professional psychological counselors to solve the problem of delayed intervention, ensuring timely follow-up from early warning to intervention to feedback, and effectively reducing the risk of psychological problems worsening due to lack of timely intervention. Attached Figure Description

[0015] Figure 1 This is a system block diagram of the intelligent early warning system for adolescent mental sub-health based on multi-source heterogeneous data fusion proposed in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0017] Please see Figure 1 As shown, this invention is an intelligent early warning system for adolescent mental sub-health based on multi-source heterogeneous data fusion, comprising: A smart early warning system for adolescent mental health sub-health based on multi-source heterogeneous data fusion includes: Data Acquisition Layer: Collects three-dimensional data on adolescent behavior, physiology, and social behavior through multiple heterogeneous channels, including campus card consumption records, physiological data monitored by wearable devices, social media language analysis, and questionnaire surveys. After data cleaning, noise reduction, and standardization, a raw data pool is constructed. Data fusion processing layer: The weighted average and Kalman filter fusion algorithms are used to integrate multi-source data. Combined with feature selection and extraction techniques, key predictive indicators of sub-psychological health are mined from the fused high-dimensional feature space to form a three-dimensional psychological profile to support model analysis. Deep learning early warning model layer: Based on the Transformer deep learning architecture, a multimodal fusion early warning model is constructed. The model is input through data preprocessing and feature engineering, trained on a large-scale labeled dataset, and the parameters are optimized through cross-validation. Ensemble learning is then used to perform real-time prediction and dynamic tracking of psychological sub-health risks. Intelligent early warning and intervention platform layer: The early warning results are displayed visually on mobile devices. It integrates a three-way linkage mechanism between the management end, the teacher end, and the parent end. The built-in AI psychological counseling module provides 24-hour online intervention, and high-risk cases are automatically referred to professional psychological counselors.

[0018] In one embodiment, constructing the original data pool includes the following steps: By connecting to the school's campus card system API, structured data such as campus card consumption records (time / location / amount), access control system entry and exit times, and library borrowing frequency are obtained. With student authorization, time-series data such as heart rate, sleep quality, and steps are synchronized through wearable device manufacturer SDKs (such as nighttime heart rate variability collected by smart bracelets). Under the premise of compliance, Python web scraping frameworks (such as Scrapy) are used to collect publicly available social media text, emojis, and interactive behaviors. After filtering out noise such as advertisements using regular expressions, the data is stored in JSON format. Traditional scales such as SCL-90 and MHT are transformed into online questionnaires. Self-reported results of depression and anxiety are collected electronically through a psychological assessment system and linked to student IDs to form a basic psychological profile. The depression risk level (low / medium / high) is labeled based on the results of psychological assessments, and the sentiment tendency (positive / neutral / negative) of social media texts is labeled using NLP models to form a structured labeling system; Tiered storage: Student Basic Information Table: Stores static information such as student ID, gender, age, grade, left-behind status, and frequency of parent-child communication, supporting subsequent group analysis and risk factor association research; Multi-source data master table: behavioral, physiological and social data are associated by timestamp, including raw data, cleaned data and annotation results, forming the core data base for dynamic tracking; Risk warning label table: Records warning signals (such as depressive tendencies), warning levels and intervention status output by the model, supporting real-time warnings and closed-loop interventions.

[0019] In one embodiment, the integration of multi-source data using a weighted average and Kalman filter fusion algorithm includes the following steps: The weighted average weighting is based on prior research or expert knowledge to assess the degree of influence of each feature on the psychological state. For example, negative sentiment on social media may be given a higher weight (e.g., 0.3), while fluctuations in consumption frequency may be given a lower weight (e.g., 0.1). The sum of the weights of all features is 1. Kalman filtering dynamic fusion is adopted to define a state vector containing psychological state indicators such as depression level and anxiety level, such as [depression score, anxiety score]. Based on the pattern of psychological state change, a state transition equation is established to describe the trend of state vector change over time. Multi-source features are mapped to the state vector, such as associating social media sentiment score with depression score. Initialize the state vector and covariance matrix to reflect the estimation uncertainty at the initial time step. Predict the current state vector and covariance matrix using the state transition matrix. Calculate the Kalman gain and adjust the predicted state by combining it with weighted average observations (such as weighted social media sentiment scores). Update the state estimate and covariance matrix to provide a basis for the prediction of the next time step, forming a dynamic closed loop.

[0020] It outputs accurate estimates of psychological state indicators such as depression level and anxiety level at the current moment, forming a dynamic psychological profile. It quantifies the uncertainty of the estimate through covariance matrix, reflects the correlation and confidence interval between various state indicators, and triggers psychological sub-health warnings (such as generating a high-risk signal when the depression score is >0.7) based on the comparison between the state estimation results and preset thresholds, providing a basis for subsequent intervention.

[0021] In one embodiment, the method of combining feature selection and extraction techniques to mine key predictive indicators of sub-optimal mental health from the fused high-dimensional feature space includes the following steps: The data from four dimensions—campus card spending records, wearable device physiological data, social media texts, and psychological assessment results—are integrated into a high-dimensional feature matrix that includes numerical (spending frequency), time-series (heart rate fluctuations), textual (emotional tendencies), and subjective (depression scores). Using individual adolescents as rows and feature dimensions as columns, a matrix is ​​constructed that includes features such as consumption patterns, changes in physiological indicators, and emotional tendencies on social media. A chi-square test is performed on each feature and its correlation with sub-health states (such as depression / anxiety) to calculate the p-value to assess the correlation between the feature and the state. For example, the frequency of use of negative emotion words on social media may show a significant correlation with depression (p<0.05), while fluctuations in consumption frequency may have a weaker correlation. A p-value threshold (such as 0.05) is set to select a subset of features that are significantly related to sub-health states as input for subsequent feature extraction. Linear transformation is performed on the filtered feature matrix to extract the top N principal components (such as principal components with a cumulative variance contribution rate > 85%), which preserves the overall data structure while reducing dimensionality. For example, principal component analysis (PCA) is used to integrate features such as consumption frequency, heart rate variability, and social media sentiment into three principal components: “behavioral stability”, “physiological state fluctuation”, and “social sentiment”. For the task of distinguishing between sub-healthy and non-sub-healthy mental states, linear discriminant analysis (LDA) is used to find the discriminative direction, maximize the inter-class variance, minimize the intra-class variance, and extract discriminative feature vectors. Features extracted from PCA, LDA, and self-attention mechanisms are integrated into three dimensions: behavior, physiology, and social. The behavior dimension includes participation in campus activities and regularity of daily routines; the physiology dimension includes heart rate variability and sleep quality; and the social dimension includes frequency of social media interactions and emotional tendencies. Based on the integrated features, a three-dimensional psychological profile containing specific values ​​or categories is generated for each adolescent. For example, the profile may show that a student has a "behavioral dimension score of 0.8 (high activity level), physiological dimension score of 0.6 (normal fluctuation), and social dimension score of 0.3 (negative tendency)", and the overall assessment is "low risk". Through heatmaps, radar charts, or dashboards, the behavioral, physiological, and social characteristics and overall risk levels can be presented intuitively. For example, radar charts can show the distribution of students' scores in each dimension, while heatmaps can show the characteristic correlation patterns at the group level.

[0022] In one embodiment, constructing a multimodal fusion early warning model based on the Transformer deep learning architecture includes the following steps: The text modality uses a pre-trained BERT model to encode social media text, generating sentiment vectors (e.g., positive / neutral / negative) to capture emotional features in social interactions. The physiological modality uses a sliding window to extract statistical features (e.g., 7-day average heart rate, deep sleep ratio) from time-series data such as heart rate and sleep quality. The behavioral modality performs frequency statistics and amount distribution analysis on campus card spending records to generate consumption pattern feature vectors (e.g., high-frequency small-amount spending tendency). The subjective modality converts psychological assessment scores into standardized values ​​to form continuous variables of depression / anxiety tendency. An independent Transformer encoder is designed for each modality (behavioral / physiological / social / subjective). Each layer contains a multi-head self-attention mechanism and a feedforward neural network to capture local and global features within the modality. For example, the social text encoder focuses on "negative emotion words" through self-attention, while the physiological data encoder captures heart rate variability trends. A cross-modal attention mechanism is used at the encoder output to dynamically weight and fuse feature vectors from different modalities. For example, the "negative emotion" feature of social text and the "heart rate variability" feature of physiological data are associated through attention weights to form a fused feature vector. We use transfer learning to initialize model weights, combine data augmentation techniques (such as text synonym replacement and physiological data noise injection) to improve generalization ability, and optimize hyperparameters such as learning rate, batch size, and number of attention heads through grid search to achieve optimal performance on the validation set. The model output includes the probability of sub-optimal mental health risk, risk level (low / medium / high), and the ranking of the importance of key features (such as 30% for negative tendencies in social texts and 25% for abnormal physiological data), supporting intervention decision-making.

[0023] In one embodiment, the real-time prediction and dynamic tracking of psychological sub-health risks includes the following steps: The preprocessed feature vectors are captured by a self-attention layer to capture cross-modal associations (such as the association between social text "negative emotions" and physiological data "heart rate variability"), and the risk probability is output through a fully connected layer (such as 0.8 indicating high risk). The model output includes the risk level (low / medium / high), probability value and contribution of key features (such as social modality accounting for 40%), and is transmitted to the early warning platform. The risk trends of individuals / groups are dynamically presented through heat maps and line graphs. For example, if a student's risk probability increases from 0.3 to 0.7 in the past 7 days, high-risk cases will automatically trigger APP push, SMS and email, and be synchronized to the management terminal, teacher terminal and parent terminal, start the "AI psychological counseling module" for 24-hour emotional support, and refer to professional psychological counselors. By continuously collecting behavioral, physiological, and social data through wearable devices and campus systems, an individual psychological state time series database is constructed, recording historical changes in state (such as fluctuations in sleep quality from the beginning to the end of the semester), generating a psychological profile report every month, and forming a dynamic risk profile by combining SCL-90 scale scores, model prediction results, and intervention effects. Kalman filtering is used to predict psychological state trends. Abnormal fluctuations (such as a sudden 20% drop in heart rate variability) are detected by the 3σ principle, triggering a primary warning. Thresholds are set according to risk probability (such as >0.6 triggering a medium warning and >0.8 triggering a high warning), and the intensity of intervention is automatically adjusted (such as increasing the frequency of AI tutoring and referring to professional counseling).

[0024] In one embodiment, the tripartite linkage mechanism between the management terminal, the teacher terminal, and the parent terminal includes: Standardized interfaces are built using RESTful API / GraphQL protocols to support the access of heterogeneous data, such as campus behavioral data (e.g., consumption records), physiological data (e.g., heart rate), and social media text (analyzed by NLP sentiment analysis), ensuring HTTPS encrypted transmission and compliance with the "Regulations on the Protection of Children's Personal Information Online". Sensitive data (such as student names) is anonymized. The management side can access the school-wide aggregated data and individual risk profiles, while the teachers side can only access the behavior and early warning information of the students in their classes, and the parents side can only access their own children's physiological and emotional reports, forming a three-level data view of macro, meso and micro. Integrate third-party messaging services and customize push content according to roles—the management side receives the school's high-risk list and intervention effect report, the teacher side receives warnings of students' emotional fluctuations and suggested measures (such as after-school counseling), and the parent side receives reminders of their children's abnormal emotions and family care guidelines (such as sleep improvement suggestions). The built-in video call function supports real-time communication between teachers and parents. Voice-to-text conversion automatically generates communication records and archives them in the student's psychological profile, ensuring that the communication content is traceable. Based on RBAC, the access control system defines three roles: administrator (system configuration / data export), teacher (intervention work order creation / class data viewing), and parent (family information update / children's report viewing). Permission changes are controlled through OA approval flow to ensure compliance. All roles' access to and operations on sensitive data (such as modification of warning thresholds) are recorded. Logs are stored on the blockchain to ensure immutability and support filtering and backtracking by time, user, and operation type. High-risk students automatically generate work orders and assign responsibility to relevant parties—for mild risks, a notification is sent to the teacher's end suggesting classroom attention; for moderate risks, a notification is sent to the management end to coordinate the intervention of a psychological counselor; for severe risks, an emergency process is triggered to notify parents and coordinate with the school doctor for on-site assessment. The work order status is updated in real time, supporting the marking of processing progress and results by all three parties. The effects of different measures are compared through A / B testing (such as teacher counseling vs. family care). Evaluation indicators include the rate of improvement in mood, changes in academic performance, and the number of re-warnings.

[0025] In one embodiment, the built-in AI-powered psychological counseling module provides 24-hour online intervention, including the following steps: By integrating voice, text, physiological signals (such as heart rate / skin conductance activity from wearable devices), and behavioral data (such as campus card spending patterns), NLP is used to extract semantic features from the text (such as BERT model to analyze emotion labels), CNN is used to process facial expression images, and LSTM is used to analyze speech spectral features (such as fundamental frequency and energy) to form a structured feature vector. The features of each modality are dynamically weighted through the Transformer self-attention mechanism to construct a comprehensive representation vector of the user's psychological state, supporting subsequent emotion computing and risk assessment. Based on a pre-trained BERT-NLP model, emotion labels (such as "anxiety -0.85" and "sadness -0.72") are identified. Combined with a keyword library (such as "suicide" and "despair"), emergency alerts are triggered. A bidirectional LSTM classifier is used to analyze features such as trembling tone and abnormal speech rate to help judge the intensity of emotions and stress status. Heart rate variability (HRV) is monitored in real time through wearable devices. When HRV decreases, stress status is indicated. Combined with behavioral data (such as nighttime consumption frequency), the risk is comprehensively assessed. The fused features are input into a deep neural network (DNN) to output the depression / anxiety risk level (low / medium / high) and intervention priority. The assessment results are dynamically updated every 15 minutes. When the risk level rises to "medium / high", the system automatically pushes an intervention pop-up to the user (such as "You look a little tired lately, would you like to talk to me?"), guiding the user into an encrypted chat interface. The user actively inputs text / voice through the "emotional tree hole" function, and the system analyzes the emotions in real time and generates a response, supporting multi-round dialogue to adjust the strategy; Layered response: For low-risk users: Push cognitive behavioral therapy (CBT) self-help tools (such as emotion diary templates) and mindfulness meditation audio to guide them in self-regulating their emotions; For medium-risk users: Initiate a structured dialogue, guide users to describe their specific concerns, provide a solution library (such as "How to deal with test anxiety?"), and generate personalized suggestions by combining knowledge graphs (such as the DSM-5 diagnostic criteria); High-risk users: Generate emergency intervention scripts (such as "You are in danger right now, I need to contact your teacher / parents, is that okay?"), and simultaneously trigger the manual referral process; It uses the Tacotron 2 speech synthesis model to generate natural speech responses, supports dialect adaptation, and uses 3D face reconstruction technology to generate virtual counselor expressions, matching the response emotions (such as a smile to comfort, a serious reminder) to enhance the interactive experience; If a user explicitly expresses suicidal intent (e.g., "I want to jump off a building"), and the emotion computing model determines "extreme anxiety" five times in a row, an emergency referral is triggered. If the user refuses all self-help intervention suggestions and the conversation lasts for more than 30 minutes without alleviating the emotion, the intervention is upgraded to manual intervention. The system pushes pop-ups to the management and teachers' terminals, displaying the user's risk level, last interaction record, and suggested referral recipients (such as school counselors). Counselors can view user data through an encrypted channel within 5 minutes and decide whether to take over. During referral, only necessary information (such as risk level and key dialogue excerpts) is shared. User identity is anonymized, in accordance with the "Regulations on the Protection of Children's Personal Information Online". AI automatically generates referral case reports, including emotional change curves and the effectiveness of intervention measures. Counselors evaluate the referral results and optimize the risk assessment model thresholds.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion, characterized in that, include: Data Acquisition Layer: Collects three-dimensional data on adolescent behavior, physiology, and social behavior through multiple heterogeneous channels, including campus card consumption records, physiological data monitored by wearable devices, social media language analysis, and questionnaire surveys. After data cleaning, noise reduction, and standardization, a raw data pool is constructed. Data fusion processing layer: The weighted average and Kalman filter fusion algorithms are used to integrate multi-source data. Combined with feature selection and extraction techniques, key predictive indicators of sub-psychological health are mined from the fused high-dimensional feature space to form a three-dimensional psychological profile to support model analysis. Deep learning early warning model layer: Based on the Transformer deep learning architecture, a multimodal fusion early warning model is constructed. The model is input through data preprocessing and feature engineering, trained on a large-scale labeled dataset, and the parameters are optimized through cross-validation. Ensemble learning is then used to perform real-time prediction and dynamic tracking of psychological sub-health risks. Intelligent early warning and intervention platform layer: The early warning results are displayed visually on mobile devices. It integrates a three-way linkage mechanism between the management end, the teacher end, and the parent end. The built-in AI psychological counseling module provides 24-hour online intervention, and high-risk cases are automatically referred to professional psychological counselors.

2. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The construction of the original data pool includes the following steps: By connecting to the school's campus card system API, structured data such as campus card consumption records, access control system entry and exit times, and library borrowing frequency are obtained. With student authorization, time-series data is synchronized through wearable devices to collect publicly available social media text, emojis, and interactive behaviors. Online questionnaires were used to collect self-assessment results of depression and anxiety electronically through a psychological assessment system. These results were then linked to student IDs to form a basic psychological profile. Depression risk levels were labeled based on the psychological assessment results, and sentiment bias was assigned to social media texts. The raw data pool uses tiered storage: Student Basic Information Table: Stores students' static information; Multi-source data master table: behavioral, physiological and social data are associated by timestamp, including raw data, cleaned data and annotation results, forming the core data base for dynamic tracking; Risk warning label table: Records the warning signals, warning levels and intervention status output by the model, supporting real-time warning and closed-loop intervention.

3. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method of integrating multi-source data using a weighted average and Kalman filter fusion algorithm includes the following steps: The weighted average weight allocation, based on previous research or expert knowledge, assesses the degree of influence of each feature on the psychological state. Kalman filtering is used for dynamic fusion to define the state vector of the psychological state index. Based on the laws of psychological state change, a state transition equation is established to describe the trend of the state vector over time, and multi-source features are mapped to the state vector. Initialize the state vector and covariance matrix to reflect the estimation uncertainty at the initial time. Predict the current state vector and covariance matrix through the state transition matrix, calculate the Kalman gain, adjust the predicted state by combining the weighted average observation value, update the state estimate and covariance matrix, and provide a basis for the prediction of the next time step. Output the estimated value of the psychological state index at the current moment to form a dynamic psychological profile. Quantify the estimation uncertainty through the covariance matrix to reflect the correlation and confidence interval between the various state indicators. Based on the comparison of the state estimation results with the preset threshold, trigger a psychological sub-health warning.

4. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The method of combining feature selection and extraction techniques to mine key predictive indicators of sub-optimal mental health from the fused high-dimensional feature space includes the following steps: The data from four dimensions—campus card spending records, wearable device physiological data, social media texts, and psychological test results—are integrated into a high-dimensional feature matrix. This matrix includes consumption patterns, changes in physiological indicators, and social media sentiment characteristics. A chi-square test is performed on each feature and its correlation with the sub-health state. The p-value is calculated to assess the correlation between the feature and the state. A p-value threshold is set to select a subset of features that are significantly related to sub-health. The selected feature matrix is ​​then linearly transformed to extract the top N principal components, thus reducing the dimensionality while preserving the overall data structure. For the task of distinguishing between sub-healthy and non-sub-healthy mental states, linear discriminant analysis (LDA) is used to find the discriminative direction, maximize the between-class variance, minimize the within-class variance, and extract discriminative feature vectors. Features extracted from principal component analysis, LDA, and self-attention mechanisms are integrated into three dimensions: behavior, physiology, and social. Based on the integrated features, a three-dimensional psychological profile containing specific values ​​or categories is generated for each adolescent. The behavioral, physiological, and social characteristics and comprehensive risk level are presented intuitively through heatmaps, radar charts, or dashboards.

5. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The construction of a multimodal fusion early warning model based on the Transformer deep learning architecture includes the following steps: The text modality encodes social media text to generate sentiment vectors and capture emotional features in social interactions. The physiological modality uses a sliding window to extract statistical features from time-series data. The behavioral modality performs frequency statistics and amount distribution analysis on campus card consumption records to generate consumption pattern feature vectors. The subjective modality converts psychological test scores into standardized values ​​to form continuous variables of depression and anxiety tendencies. An independent Transformer encoder is designed for each modality. Each layer contains a multi-head self-attention mechanism and a feedforward neural network to capture local and global features within the modality. A cross-modal attention mechanism is used at the encoder output to dynamically weight and fuse feature vectors from different modalities. The model output includes the probability of psychological sub-health risk, risk level, and ranking of the importance of key features, supporting intervention decision-making.

6. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The real-time prediction and dynamic tracking of psychological sub-health risks includes the following steps: The preprocessed feature vectors are captured by a self-attention layer to capture cross-modal correlations, and the risk probability is output through a fully connected layer. The model output includes risk level, probability value, and contribution of key features, and is transmitted to the early warning platform. Individual and group risk trends are dynamically presented through heatmaps and line graphs. Behavioral-physiological-social data are continuously collected through wearable devices and campus systems to construct an individual psychological state time series database, record historical state changes, and generate a psychological profile report every month. Combined with SCL-90 scale scores, model prediction results, and intervention effects, a dynamic risk profile is formed. Kalman filtering is used to predict psychological state trends, detect abnormal fluctuations, trigger primary early warnings, and set thresholds based on risk probabilities.

7. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The aforementioned tripartite linkage mechanism involving the management terminal, teacher terminal, and parent terminal includes the following steps: Standardized interfaces are built to support the access of heterogeneous data such as campus behavior data, physiological characteristics data, and social media texts. Sensitive data is anonymized. The management side can access the school-wide aggregated data and individual risk profiles, while the teacher side is limited to the behavior and early warning information of students in their own class, and the parent side is limited to the physiological and emotional reports of their own children. Integrate third-party messaging services and customize push content according to roles: the management side receives the school's high-risk list and intervention effect report, the teacher side receives the class's student emotional fluctuation warning and suggested measures, and the parent side receives the child's abnormal emotional reminder and family care guide; The built-in video call function supports real-time communication between teachers and parents. Voice-to-text communication records are automatically generated and archived in the student's psychological profile. Three roles are defined: administrator, teacher, and parent. All roles' access to and operation of sensitive data are recorded. High-risk students automatically generate work orders and assign responsibility: mild risk is pushed to the teacher's end to suggest classroom attention, moderate risk is pushed to the management end to coordinate the intervention of psychological counselors, and severe risk triggers the emergency process to notify parents and coordinate with the school doctor for on-site assessment. The work order status is updated in real time and supports the three parties to mark the processing progress and results. The effects of different measures are compared through A / B testing. The evaluation indicators include the rate of improvement in mood, changes in academic performance, and the number of re-warnings.

8. The intelligent early warning system for adolescent sub-health based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The built-in AI-powered psychological counseling module provides 24 / 7 online intervention, including the following steps: By integrating voice, text, physiological signals, and behavioral data, a structured feature vector is formed. The features of each modality are dynamically weighted through the Transformer self-attention mechanism to construct a comprehensive representation vector of the user's psychological state. Identify emotion tags, trigger emergency alerts by combining them with a keyword database, analyze abnormal features, assist in judging the intensity of emotions and stress levels, monitor heart rate variability (HRV) in real time through wearable devices, indicate stress levels when HRV decreases, comprehensively assess risks by combining behavioral data, input fused features into a deep neural network (DNN), and output the risk level of depression and anxiety and the priority of intervention. When the risk level rises to medium or high, the system automatically pushes an intervention pop-up to the user, guiding the user into an encrypted chat interface. The user can actively input text and voice through the emotion tree hole function, and the system will analyze emotions and generate responses in real time, supporting multi-round dialogue to adjust strategies. Layered response: For low-risk users: Push cognitive behavioral therapy self-help tools and mindfulness meditation audio to guide them in self-regulating their emotions; For medium-risk users: initiate a structured dialogue, guide the user to describe their specific concerns, provide a solution library, and generate personalized suggestions by combining knowledge graphs; High-risk users: Generate emergency intervention scripts and simultaneously trigger the manual referral process; Generate natural voice responses, generate virtual tutor expressions, match response emotions, and enhance the interactive experience; If a user explicitly expresses suicidal intent, and the emotion computing model continuously determines that the user is extremely anxious, an emergency referral is triggered. If the user refuses all self-help intervention suggestions and the conversation lasts for more than 30 minutes without relieving the emotion, the intervention is upgraded to manual intervention. The system pushes pop-ups to the management and teachers' terminals, displaying the user's risk level, last interaction record, and suggested referral targets. Psychological counselors can view user data through an encrypted channel and decide whether to take over. Only necessary information is shared during referrals, and user identities are anonymized. AI automatically generates referral case reports, including emotional change curves and the effectiveness of intervention measures. Psychological counselors evaluate the referral results and optimize the risk assessment model thresholds.

Images