Psychological risk early warning intervention system and method based on multi-source heterogeneous portraits and agents

Abstract

The invention discloses a psychological risk early warning intervention system and method based on a multi-source heterogeneous portrait and an intelligent agent. The system comprises a multi-source dynamic psychological health portrait construction module, a psychological problem dynamic early warning module, a classification and grading intervention strategy matching module and a psychological health intelligent agent service module. The core process of the method comprises the following steps: firstly, collecting multi-dimensional dynamic data from a multi-heterogeneous data source, and constructing a real-time updated dynamic psychological health portrait after processing and fusing the multi-dimensional dynamic data; multi-dimensional psychological risk assessment and graded early warning are realized on the basis of portrayal data in combination with technologies such as machine learning and natural language processing; then, a personalized intervention scheme is matched through a classification and grading model; and finally, performing intervention by combining a retrieval enhancement generation technology and an agent of a large language model with a human-skill cooperation mode, and optimizing the system through feedback data. The method realizes the crossing of the psychological state from static evaluation to dynamic perception, improves the early warning accuracy and intervention precision, effectively makes up for the defect of professional resource deficiency, and can be widely applied to colleges, communities, enterprises and other scenes.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and mental health technology, specifically relating to a psychological risk early warning and intervention system and method based on multi-source heterogeneous profiling and intelligent agents. It is particularly suitable for the mental health management of college students and can also be extended to psychological risk prevention and control in communities, enterprises and other scenarios. Background Technology

[0002] With increasing societal attention to mental health, the mental health of college students has become a key focus in the education sector. Currently, vocational colleges and other higher education institutions face numerous challenges in their work on student mental health: Data is often static and limited in scope, relying primarily on traditional psychological scales with long update cycles, resulting in delays and an inability to reflect the dynamic changes in students' mental states in real time; early warning mechanisms are inaccurate, with existing warnings largely based on single indicators or manual observation, lacking multi-dimensional data correlation analysis, making it difficult to identify potential psychological risks early and accurately, leading to missed or false reports; intervention strategies are often "one-size-fits-all," lacking personalization and failing to accurately match interventions to the type and severity of students' mental health problems, resulting in wasted resources and poor intervention outcomes; professional resources are severely lacking, with an insufficient number of full-time mental health teachers, making it difficult to cover all students, especially the "silent group," and resulting in a passive service model with limited coverage.

[0003] To address these issues, research has been conducted on the application of artificial intelligence in the field of mental health. For example, some patents have explored the application of multi-source data fusion in mental health assessment, combining facial expression images, historical text emotional data, or physiological data collected by wearable devices to assess or warn of mental health status through deep learning models; other patents utilize historical data from psychological counseling to train generative dialogue models, combining wearable device data and facial expression information to monitor user emotions and generate mental health reports.

[0004] For example, in the field of multi-source data fusion, the publicly disclosed patent CN119252487A of Newland Digital Technology Co., Ltd. combines facial expression images with historical text emotional data, and uses a residual neural network model (ResNet) and a deep bidirectional language representation model (Bidirectional Encoder Representation of Transformer, BERT) to assess mental health status. The publicly disclosed patent CN119908727A of Zhejiang Oriental Vocational and Technical College uses a wearable device that integrates a heart rate monitor, skin resistance sensor and biometric sensor to collect users' physiological data such as heart rate and respiratory rate, as well as behavioral data such as sleep patterns and exercise volume. After using principal component analysis (PCA) to perform feature dimensionality reduction, the feature set is converted into a mental state image, and then a deep learning model is used to provide early warning of mental illness. In the area of ​​crisis signal detection, Newland Digital Technology Co., Ltd.'s published patent CN119252486A utilizes historical data from psychological counseling to train a generative dialogue model based on the Continuous Bag of Words (CBOW) and Long Short-Term Memory (LSTM) networks. During the dialogue interaction, it monitors the counseling content in real time, along with information such as heart rate collected by wearable devices and facial expressions captured by cameras. The neural network model monitors the user's emotions in real time and generates a mental health report. These patented technologies, through the fusion analysis and real-time computation of multimodal data, provide an intelligent solution for mental health diagnosis, demonstrating the application value of artificial intelligence technology in the early identification and intervention of mental illnesses.

[0005] However, existing technologies still have significant shortcomings: on the one hand, the multidimensional data relied upon by current research (such as physiological data from wearable devices and social network data) is difficult to collect in university information systems, limiting the operability of research results in practical application scenarios; on the other hand, traditional chatbots are not specifically designed for psychological counseling, their dialogue logic and response strategies lack professionalism, and may even exacerbate students' psychological dependence or trigger negative emotions such as depression due to improper guidance, thus having a counterproductive effect on students' mental health. Therefore, there is an urgent need for an intelligent solution that can integrate multi-source data, achieve dynamic early warning, provide precise intervention, and compensate for insufficient manpower. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to overcome the shortcomings of the prior art and provide a psychological risk early warning and intervention system and method based on multi-source heterogeneous profiling and intelligent agents, so as to realize dynamic perception, accurate early warning and personalized intervention of psychological state, build a "human-technology collaboration" service model, make up for the lack of professional resources, and improve the timeliness, accuracy and sustainability of mental health services.

[0007] Technical solution: The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents described in this invention includes:

[0008] A multi-source dynamic mental health profile construction module is used to collect multi-source heterogeneous data, process and fuse it, and generate a dynamic mental health profile that is updated in real time. The multi-source dynamic mental health profile construction module includes a data acquisition unit, a data processing and fusion unit, and a dynamic profile generation unit.

[0009] The dynamic early warning module for mental health issues is connected to the multi-source dynamic mental health profile construction module. It is used to monitor changes in key indicators of the profile in real time and combine text data analysis to achieve multi-dimensional risk assessment and graded early warning. The dynamic early warning module for mental health issues includes a real-time monitoring unit, a multi-dimensional risk assessment unit, a natural language processing and recognition unit, and an early warning triggering unit.

[0010] The classification and grading intervention strategy matching module is connected to the dynamic early warning module for psychological problems. It is used to classify and grade the psychological problems identified in the early warning and match them with personalized intervention plans. The classification and grading intervention strategy matching module includes a classification and grading model unit, a differentiated intervention plan library, and an intelligent matching unit.

[0011] The mental health intelligent agent service module is connected to the classification and grading intervention strategy matching module, and is used to execute personalized intervention plans and realize human-technology collaborative task distribution; the mental health intelligent agent service module includes an AI diagnostic expert core, a multimodal interaction unit, and a human-technology collaborative task distribution unit;

[0012] The modules work together to achieve full-cycle management of psychological risks.

[0013] In some implementations, the data acquisition unit is used to collect various multi-dimensional dynamic data from university, community, and enterprise systems. For example, in a university setting, it can collect academic data, behavioral data, online behavioral data, social activity data, classroom performance data, and psychological assessment data from academic affairs offices, student affairs offices, campus card systems, online learning platforms, psychological assessment systems, and social platforms / campus activity systems. Academic data includes grades, course selection, absence records, and failing records; behavioral data includes consumption frequency, consumption location, library entry / exit records, and dormitory entry / exit records; online behavioral data includes login duration, study frequency, and assignment submission status; social activity data includes club participation, activity registration, and social media interaction; classroom performance data includes classroom status and classroom interaction activity; and psychological assessment data includes psychological assessment records and scores on various scales. By covering multiple data sources, the comprehensiveness and diversity of the collected data are ensured, providing rich data support for subsequent profile construction.

[0014] In some implementations, the data processing and fusion unit employs feature engineering and data mining techniques such as surveys, averaging, and K-nearest neighbor imputation to clean and denoise the collected multi-source data, addressing issues of data inconsistency and missing values. It uses techniques like Z-scores to standardize data of different dimensions, eliminating the impact of dimensional differences on subsequent analysis. Furthermore, it breaks down departmental data barriers, establishing a unified data association model based on unique identifiers of object information (student names, student IDs, etc.) to effectively integrate multi-source data. This unit improves data quality through standardized data processing procedures, laying the foundation for data fusion and profile generation.

[0015] In some implementations, the dynamic profile generation unit constructs a multi-dimensional dynamic profile model of an object's mental health based on processed data, using techniques such as dynamic graph association models. This model is continuously updated using machine learning algorithms, dynamically analyzing information across various dimensions. Subsequently, interpretable machine learning techniques are used to uncover a set of key indicators affecting mental health, thereby generating a real-time updated digital profile of mental health for each object. The specific process includes the following:

[0016] ① A multi-dimensional dynamic association graph model is constructed using temporal graph neural networks (DGNN) or dynamic knowledge graph embedding technology. Each object is defined as a central node, and its various data such as grades, consumption, and social interactions are encoded into the multi-dimensional features of the node. Association edges such as "academic-consumption association" and "social-activity association" are established between different dimensions, with the edge weights dynamically learned by the model rather than pre-set. The model receives data streams over continuous time periods, automatically learns and updates the strength of associations between node features and dimensions, and finally outputs a dynamically updated personalized multi-dimensional association matrix, quantifying the real-time relationship strength of various indicators of the object and revealing potential behavioral pattern combinations.

[0017] ② The personalized key indicator set mining engine is based on surrogate prediction models such as Lightweight Gradient Boosting Tree (LightGBM) to predict the trend of the psychological state of the target. When working, the multi-dimensional feature vector of the target is first input into the surrogate model, and then the interpretability engine such as SHAP is used for reverse analysis to calculate the contribution value of each input feature to the model output (such as SHAP value). Then, the features are sorted according to the absolute value of the contribution value and the top N features are selected to form a personalized key indicator set that affects the predicted psychological state of the target. The key indicators are different for different targets. For example, the key indicators for target A may be "frequency of dining in the cafeteria" and "on-time attendance rate in the morning", while those for target B may be "participation in online group discussions" and "number of times of use of sports venues".

[0018] In the overall workflow, after receiving the multi-dimensional time-series feature vector of the object, the unit first generates / updates the individual multi-dimensional association matrix through the dynamic graph association model, and then outputs a personalized list of key indicators and contribution with the help of the interpretability engine. Finally, the original feature snapshot, association insight, key indicator set, and profile metadata are encapsulated into a dynamic profile data object and pushed to the data warehouse or message queue to provide core input for the risk level determination and early warning decision of the subsequent risk assessment unit.

[0019] In some implementations, the real-time monitoring unit is connected to the dynamic profile generation unit of the profile construction module. This unit monitors the dynamic changes of personalized key indicators in the subject's mental health profile in real time, and promptly triggers subsequent risk assessment processes when abnormal fluctuations occur in the indicators. Through continuous monitoring, this unit ensures that potential changes in the subject's mental state are detected as soon as possible.

[0020] In some implementations, the multi-dimensional risk assessment unit utilizes a constructed psychological problem early warning model to comprehensively analyze multi-dimensional information such as academic pressure, social isolation, and behavioral abnormalities, generating a comprehensive risk score. As the core of the system's decision-making, the multi-dimensional risk assessment unit employs a hierarchical fusion assessment model to transform the key indicator set and correlation matrix summary from the dynamic profile generation unit into a quantitative comprehensive risk score. This model can be divided into two layers: the first layer calculates dimensional risk factors, receiving a personalized key indicator set with SHAP contribution values, and then calculating independent risk scores for preset dimensions such as academic pressure and social isolation using a built-in lightweight sub-neural network or weighted fusion formula, combined with weights initialized and fine-tuned by expert experience; the second layer performs comprehensive risk fusion and scoring, using the risk factor scores of each dimension and cross-dimensional synergistic effect signals in the correlation matrix as input, and learning complex interaction relationships between dimensions using a gradient boosting decision tree (LightGBM) or deep residual network fusion processor, outputting a comprehensive risk probability score in the 0-1 range, while simultaneously generating a risk attribution summary. Furthermore, the model periodically performs online calibration using historical false positives and false negatives to ensure the accuracy of scoring for different groups and time periods.

[0021] The Natural Language Processing (NLP) recognition unit utilizes NLP technology to automatically analyze text data related to the target (such as consultation records and social media posts) to identify sensitive content and potential risks, such as suicidal tendencies and depressive moods. Specifically, it can rely on multi-task pre-trained language models and context-enhanced analysis techniques to extract risk signals from unstructured text. For example, its base model is one or more combinations of MentalBERT, PsyBERT, or customized RoBERTa architecture models finely tuned on a mental health corpus, possessing high semantic sensitivity. This unit operates through a three-layer recognition pipeline: a high-risk content recognition layer outputs risk levels through sequence classification tasks, combined with keyword trigger lists and a rule engine for dual verification to avoid missing critical content; a psychological emotion state analysis layer analyzes text sentiment polarity and emotional vocabulary density through multi-label classification or emotion intensity regression to identify persistent emotions such as depression; and a context and risk topic extraction layer uses named entity recognition and topic clustering techniques to locate risk triggers such as academic setbacks. Finally, the unit outputs a structured NLP risk report containing high-risk signal tags, dominant emotions, and risk topics, which is then correlated with dynamic profile data for use by the early warning unit.

[0022] The warning triggering unit automatically triggers a tiered warning when the comprehensive risk score exceeds a preset threshold or the NLP unit identifies a high-risk signal, and notifies relevant management personnel via system interface, SMS, or email. Specifically, it employs a collaborative decision-making model combining a rule engine and a Bayesian network. Based on the comprehensive risk score from the risk assessment unit, the risk attribution summary, and the risk report from the NLP unit, it achieves tiered warning triggering and escalation judgment for the psychological risk of the target. The rule engine presets clear warning triggering conditions and completes the initial warning judgment based on the comprehensive risk score threshold, such as a red warning triggered when the comprehensive risk score exceeds 0.9, and an orange warning triggered when the score exceeds 0.7 and the NLP contains a high-risk signal. The Bayesian network handles signal uncertainty and conflict, assisting in determining whether to escalate the warning. The specific judgment rules are as follows:

[0023] By constructing a probabilistic graphical model that includes evidence nodes (comprehensive risk score level, NLP high-risk signal confidence level, risk attribution type), intermediate inference nodes (risk signal reliability, risk persistence), and target nodes (necessity of warning escalation), the model quantifies the causal relationships between nodes based on conditional probability tables. This effectively handles the uncertainty and conflict of risk signals. Through probabilistic inference, it outputs the probability distribution of the necessity for warning escalation, assisting in revising the initial warning decision and ultimately determining whether to escalate the warning level. For example, the rule engine initially judges the comprehensive risk score to be 0.65 (<0.7, orange threshold), but the Bayesian network, based on the input score level = medium, NLP high-risk signal confidence level = 0.9 (high confidence), and risk attribution = family factors (long-term), infers a 75% probability that the necessity for warning escalation = "must be escalated." Ultimately, the decision exceeds the rule threshold, and the warning is upgraded to orange.

[0024] The alert system is divided into three levels: red, orange, and yellow, each corresponding to a different notification strategy. For example, a red alert notifies counselors and school counselors in real time through multiple channels, simultaneously synchronizing core risk information. An orange alert generates a work order to be completed within 24 hours, and a yellow alert is updated on the counselor's dashboard and included in regular monitoring. Simultaneously, the unit supports alert handling feedback, with feedback data flowing back to the risk assessment model and alert rule base, forming an intelligent closed loop of "alert-handling-feedback-optimization" to achieve continuous calibration of the model and rules.

[0025] In some implementations, the classification and grading model unit trains a machine learning classifier based on the object's psychological profile data and historical intervention cases to establish a psychological problem classification and grading model. This model automatically classifies and grades the object's psychological problems according to type and severity. This unit uses a multi-task hierarchical classification model as its core technology, transforming risk assessment results and dynamic profile data into actionable labels for psychological problem types and severity levels.

[0026] The system first performs input feature engineering, integrating key indicator sets from the dynamic profiling unit, risk factors and comprehensive scores from various dimensions of the risk assessment unit, risk topics and sentiment labels from the NLP unit, and historical sequence information on the object's past classification and grading. Then, it completes grading and classification through a two-layer model: the first layer uses models such as ordered logistic regression or neural networks to determine ordered severity levels such as "normal concern" and "mild distress" based on the quantitative mapping of the comprehensive risk score and the quantity and extreme degree of key indicators; the second layer uses deep neural networks or tree models for multi-label classification, combining basic features with the association patterns in the dynamic profiling to identify problem types such as "academic anxiety" and "social withdrawal." The two-layer model can construct a multi-task joint training architecture sharing underlying features to ensure consistency of results, ultimately outputting structured labels containing severity level, main type, and confidence level. The model uses a desensitized historical intervention case library as training data, and simultaneously relies on an active learning mechanism to mark low-confidence cases as expert review items, using expert judgment results to feed back into the model iteration.

[0027] The differentiated intervention program library stores standardized intervention programs designed for different problem types and severity levels. This library is a dynamically computable knowledge base, with core technologies including knowledge graph representation and causal inference analysis. In program construction and management, each intervention program is deconstructed into standardized, composable operation nodes and stored in a graph database. These nodes encompass intervention actions such as "counselor counseling" and "mindfulness group counseling," machine-readable applicability conditions such as "suitable for mild / moderate academic anxiety," as well as pre- and post-dependencies between actions and program metadata. At the dynamic optimization level, the program library is bidirectionally linked to a historical intervention case library. Through causal inference analysis based on propensity score matching, the actual effectiveness of the programs on similar individuals is evaluated. This data-driven approach optimizes the applicability conditions and recommendation weights of the programs, enabling continuous iterative improvement.

[0028] In some implementations, the intelligent matching unit automatically matches and recommends the most suitable personalized intervention plan for the object from the intervention plan library based on the output of the classification and grading model. A combination of rule-based initial screening, collaborative filtering, and similarity calculation ranking techniques is used to achieve accurate matching between intervention plans and objects. First, through rule-based graph querying, matching candidate plans are initially screened from the intervention plan knowledge graph based on the problem type and severity level output by the classification and grading model. Then, the cosine similarity or Euclidean distance of the dynamic profile features of the current case object and historical case objects is calculated, and combined with the intervention effect data of similar objects, a matching intervention plan is recommended.

[0029] In some implementations, the core of the AI ​​diagnostic expert integrates retrieval-enhanced generation (RAG) technology and a large language model (LLM), and accesses mental health profile data, early warning signals, and a mental health knowledge graph containing psychological knowledge and typical cases. This enables it to understand complex psychological problems, conduct preliminary psychological diagnoses, and generate professional dialogue content and intervention suggestions. The AI ​​diagnostic expert uses a core architecture of "retrieval-enhanced generation (RAG) + agent," integrating large language model (LLM) capabilities. Knowledge retrieval enhancement and security constraints ensure the professional reliability of the diagnosis and suggestions. Its workflow begins with context awareness and problem understanding. The agent first integrates multi-source data such as the object's dynamic profile, risk assessment results, classification and grading labels, and interactive dialogue history. Then, a dedicated lightweight Transformer encoder compresses this information into structured "contextual cues," providing accurate context for the LLM. Subsequently, it enters the knowledge retrieval and enhancement stage, performing vector retrieval, graph relationship retrieval, and case similarity retrieval based on the object's problem, obtaining authoritative support from the mental health knowledge graph and a desensitized typical case library. In the diagnostic reasoning and content generation stages, LLM uses a "thinking chain" format to reason step by step, combining the role boundaries and responsibilities defined by the system prompts to generate output that conforms to the predefined template. Key assertions must be verified by a fact checker and compared with retrieved knowledge. Finally, by combining intervention knowledge and recommendation schemes, specific personalized professional advice is generated, ensuring that all content originates from authoritative knowledge and eliminating illusions and unprofessional expressions.

[0030] In some implementations, the multimodal interaction unit supports natural interaction with objects through various means such as text and voice, and has functions such as intelligent chat and virtual social scenario training, enabling it to proactively capture the potential psychological needs of the object. This unit covers three major channels: voice, text, and vision. The voice channel uses an ASR model that integrates emotion recognition to output voice emotion tags and paralinguistic features during transcription, while TTS technology can adjust the tone to adapt to the dialogue content. The text channel uses fine-grained emotion analysis and semantic analysis to understand deeper needs. The visual channel (such as in video consultation scenarios) uses facial motion unit analysis to capture micro-expressions and combines posture recognition to judge body language status. At the dialogue management level, the system tracks and maintains dynamic conversation information such as topics and emotions through dialogue status tracking. The dialogue strategy engine can execute strategies such as proactive care (such as initiating inquiries when the object's profile deteriorates), Socratic questioning (guiding self-exploration), and crisis transfer (switching to intervention scripts when high-risk signals are detected). At the same time, the unit has a virtual social scenario training function, which constructs social anxiety scenarios through text-to-scenario generation technology, analyzes the interaction between the object and the virtual character, and provides feedback from the dimensions of social skills and emotion regulation. Ultimately, the unit generates an "instant interaction summary" from the multimodal signals in the interaction and feeds it back to the AI ​​diagnostic expert core and the dynamic profile generation unit in real time, forming an enhanced closed loop of "interaction-perception-cognition-re-interaction".

[0031] In some implementations, the human-technology collaborative task distribution unit adopts an "offline + online" integrated service model, forming a collaborative team composed of "dedicated psychological teachers + intelligent agents + counselors," which automatically assigns tasks based on the warning level and problem type. The intelligent agent is responsible for 24 / 7 online daily consultations, guidance for mild psychological problems, preliminary screening and identification of complex problems, and proactively reaching out to the "silent group" to achieve universal basic service coverage. Dedicated psychological teachers are responsible for handling severe psychological crisis intervention tasks assigned by the system (such as severe crisis cases corresponding to red warnings), conducting professional crisis interviews, and developing and implementing intervention plans. Counselors receive system instructions and assist in conducting basic psychological support and follow-up work (such as regular counseling for orange warning cases and tracking the implementation of intervention plans). Through clear role division and task allocation, professional resources are optimally allocated, improving intervention efficiency.

[0032] On the other hand, the present invention also discloses a psychological risk early warning and intervention method for the above-mentioned system, comprising the following steps:

[0033] S1. Multi-source Data Acquisition and Preprocessing: The system collects multiple data sources from universities, communities, and enterprises through its data acquisition unit. For example, in a university setting, it can automatically and continuously acquire diverse data such as academic performance, campus spending, social activities, and psychological assessment results from multiple heterogeneous data sources, including the academic affairs office, student affairs office, and campus card system. In this step, the system uses feature engineering and data mining techniques such as surveys, averaging, and K-nearest neighbor imputation to clean and denoise the raw data. Z-score technology is used for standardization to address inconsistencies and missing values. Simultaneously, the data acquisition unit directly transmits unstructured data such as consultation records and social media texts to the natural language processing unit of the early warning module, preparing it for subsequent analysis.

[0034] S2. Dynamic Mental Health Profile Construction: The data processing and fusion unit integrates the cleaned multi-source data. The dynamic profile generation unit uses temporal graph neural networks (DGNN) or dynamic knowledge graph embedding technology to automatically learn and update the correlation strength between node features and dimensions, constructing a digital profile that reflects the overall mental state of the subject. This profile is continuously updated through online learning technology, tracking the dynamic changes in the subject's mental state in real time. Simultaneously, the personalized key indicator set mining engine, based on lightweight gradient boosting trees and other surrogate prediction models, selects a set of personalized key indicators that affect the predicted mental health status of the subject, providing accurate data support for subsequent risk assessment.

[0035] S3. Psychological Risk Identification and Early Warning: The real-time monitoring unit in the early warning module continuously monitors key indicators in the profile. The multi-dimensional risk assessment unit calls a hierarchical fusion assessment model to comprehensively analyze multi-dimensional information such as academic and social aspects, calculating a comprehensive risk score and risk attribution summary. Simultaneously, the parallel Natural Language Processing (NLP) unit uses pre-trained models such as MentalBERT, PsyBERT, or a customized RoBERTa architecture model to perform sentiment analysis and sensitive content identification on the previously segmented text data, outputting a structured NLP risk report containing high-risk signal tags, dominant emotions, and risk themes. The early warning triggering unit integrates all the above analysis results and adopts a collaborative decision-making model of rule engine + Bayesian network. Once the risk exceeds a preset threshold, it automatically generates a graded early warning signal and notifies relevant personnel according to the corresponding strategy.

[0036] S4. Intervention Program Matching: Upon the generation of an early warning signal, the classification and grading intervention strategy matching module responds immediately. The classification and grading model unit first accurately identifies and grades the psychological problems indicated by the early warning, outputting structured labels containing severity level, main type, and confidence level. Subsequently, based on this judgment, the intelligent matching unit uses a combination of rule-based initial screening, collaborative filtering, and similarity calculation ranking techniques to retrieve and match the 1-3 most suitable personalized intervention programs from the differentiated intervention program library, achieving precise planning for "one person, one policy".

[0037] S5. Intervention Implementation: The matched intervention plan is sent to the mental health intelligent agent service module. This module's AI diagnostic expert core uses a "Retrieval Enhancement Generation (RAG) + Agent" architecture, integrating large language model capabilities to generate professional diagnostic content and intervention suggestions. Through a multimodal interaction unit, it engages in natural dialogue with the recipient via text and voice, performing online consultations, guidance, and virtual social scenario training. Simultaneously, the human-technology collaborative task distribution unit intelligently assigns tasks to different roles based on the severity of the problem: the AI ​​agent handles routine consultations and mild issues; counselors provide basic support and follow-up; and dedicated mental health teachers intervene to handle severe psychological crises, forming a highly efficient and collaborative intervention implementation network.

[0038] S6. Feedback and Optimization Closed Loop: All detailed processes of intervention execution, feedback information from the subjects, and effect evaluation data (such as post-intervention risk scores and satisfaction scores) are used as new data sources and reintegrated into the system's data acquisition unit. This feedback data is used for incremental training of the model and online learning of parameters, continuously optimizing the accuracy of the early warning model, the precision of the classification and grading model, and the interaction and diagnostic capabilities of the agent, enabling the entire system to continuously iterate and achieve continuous improvement in service effectiveness.

[0039] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) Achieving a leap from static assessment to dynamic perception: By integrating multi-source dynamic data to construct a real-time updated profile of the subject, the passive and lagging situation of traditional scale measurement has been reversed. It can proactively and continuously monitor the dynamic changes in the psychological state of the subject, laying a solid data foundation for accurate identification and intervention.

[0041] (2) Improved the timeliness and accuracy of early warning: Combining multidimensional profiling analysis and NLP technology, it breaks through the limitations of single indicators and manual screening, and can discover potential psychological crises earlier and more accurately, output risk attribution summaries to clarify risk triggers, effectively overcome the lag and subjectivity of traditional manual screening, and firmly grasp the "golden window period" for psychological intervention.

[0042] (3) It achieves precision and personalization of intervention: The psychological problems are automatically classified and graded by the classification and grading model, and the most suitable personalized intervention strategy is accurately matched from the solution library by the intelligent matching algorithm. The "one-size-fits-all" extensive intervention mode is abandoned, which significantly improves the pertinence of intervention and the efficiency of resource utilization, so that mental health services can truly meet the unique needs of each individual.

[0043] (4) A new service model of "human-technology collaboration" has been constructed: AI has been upgraded from "assistive tool" to "expert-like intelligent agent" that can independently undertake basic work, forming a hierarchical linkage work network with full-time teachers and counselors; the intelligent agent achieves 24 / 7 universal service coverage and proactive outreach to "silent groups", while human experts focus on complex crisis management, effectively solving the core pain points of weak professional teaching staff and limited coverage, and constructing a comprehensive mental health support network.

[0044] (5) Possesses continuous optimization and self-evolution capabilities: Through a complete data loop of “monitoring-early warning-intervention-tracking-feedback-optimization”, the intervention effect data is continuously fed back to drive model iteration, enabling the system to continuously adapt to the changes in the psychological characteristics of different groups and improve the sustainability and applicability of services.

[0045] (6) Wide range of application scenarios: This invention is not only applicable to university scenarios, but can also be extended to communities, enterprises and other scenarios to provide early warning and intervention for psychological problems of community residents, enterprise employees and other groups, and has strong promotional value. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the module architecture and workflow of the system in an embodiment of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0048] In the description of this invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "inner", "outer", etc., indicate the orientation or positional relationship shown, and are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0049] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0050] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0051] like Figure 1 As shown, this embodiment provides a psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents, applied to the psychological health management of 3200 freshmen in the 2025 cohort of a vocational college. The system is deployed on a private cloud server on campus, configured with an Intel Xeon Gold 6330 CPU, 128GB of memory, and an NVIDIA A100 GPU. It uses Python 3.9 as the development language, TensorFlow 2.10 as the deep learning framework, MySQL 8.0 for storing structured data, and Neo4j as the knowledge graph database. The system includes a multi-source dynamic psychological health profiling construction module, a dynamic early warning module for psychological problems, a classification and grading intervention strategy matching module, and a psychological health intelligent agent service module. The data input layer integrates six core data sources within the campus, while the execution and feedback layer consists of a collaborative team of 20 counselors, 3 full-time psychological teachers, and an AI intelligent agent, achieving efficient execution and feedback loops for intervention tasks.

[0052] In this embodiment, the data acquisition unit automatically collects multi-source data through the campus data platform interface. The acquisition cycle is set as follows: academic data and psychological assessment data are incrementally synchronized once a day; behavioral data and online behavior data are synchronized once an hour; and social activity data and classroom performance data are synchronized in real time. Specific data collection includes: collecting students' mid-term grades (including individual subject scores, average scores, and rankings), course selection lists, absence records (accurate to the course period), and failing records from the Academic Affairs Office system; collecting student scholarship, grant, and loan information, disciplinary records, and dormitory evaluation results from the Student Affairs Office system; collecting October campus card consumption data (including consumption time, location, and amount, with consumption locations categorized as canteen, supermarket, convenience store, print shop, etc.), library entry and exit records (including entry time, exit time, and duration of stay), and dormitory entry and exit records (including early departure time and late return time) from the campus card system; and collecting students' weekly login time (average daily time) from the online learning platform (Chaoxing Learning Platform). Login duration ≥ 2 hours is counted as high frequency), learning frequency (daily average number of clicks on learning resources ≥ 5 times is counted as active), and homework submission status (late submission, missing submission, and excellent homework records); scores of each dimension of the 2025 freshman psychological screening SCL-90 scale (9 dimensions including somatization, obsessive-compulsive symptoms, interpersonal sensitivity, etc., with a score ≥ 2 points being counted as abnormal) and UPI scale screening results are collected from the psychological assessment system; student club participation lists (including participation duration and position) and registration and sign-in records for October campus activities (such as freshman orientation lectures and sports events) are collected from the campus social platform (WeChat Moments) and publicly posted text content by students (after filtering out privacy information) are collected from the campus social platform (WeChat Moments).

[0053] The data processing and fusion unit employs the following specific process to handle data: First, data cleaning is performed, removing abnormal data with amounts ≤0.01 yuan from campus card spending and invalid data with login durations ≤10 seconds on the learning platform. K-nearest neighbor imputation technology is used to fill in the missing spending records of 32 students within the past 3 days (using the average spending of students in the same dormitory and major during the same period as a reference). Next, data standardization is performed, using Z-score technology to standardize data of different dimensions, such as grades (out of 100), spending amounts (range 0-500 yuan / week), and login durations (range 0-10 hours / week), converting them into standard scores with a mean of 0 and a standard deviation of 1. Finally, data fusion is performed, establishing a unified data association model based on student ID numbers to integrate multi-source data for the same student into structured data entries. Each entry contains 28 feature fields and is stored in a MySQL database. A data quality report is generated simultaneously to ensure data integrity ≥98% and accuracy ≥99%.

[0054] The dynamic profile generation unit uses a temporal graph neural network (DGNN) to construct a dynamic relational graph model. The model input is a continuous data stream from October 1st to October 31st, 2025, with a time step of 1 day. The node feature dimension is 28, and the relational edges are set to 8 categories (academic-consumption, academic-social, consumption-dormitory, consumption-library, social-classroom, network-academic, network-social, and psychology-academic). The model is trained using the Adam optimizer with a learning rate of 0.001 and 500 iterations. After convergence, the model automatically learns and updates the node features and relational edge weights. For example, the weight of the "academic-consumption relational" edge for student A is 0.72 (indicating a strong relation between the two), while the weight of the "social-classroom relational" edge is 0.35 (indicating a weak relation). Meanwhile, the personalized key indicator set mining engine uses Lightweight Gradient Boosting Tree (LightGBM) as a surrogate prediction model. The model input is a 28-dimensional feature vector, and the output is a psychological risk trend prediction value (0-1 interval). The SHAP interpretability engine calculates the SHAP value of each feature and selects features with an absolute SHAP value ≥ 0.1 as key indicators. The key indicators for Student A (a freshman majoring in Computer Science) were ultimately determined to be "frequency of dining in the cafeteria" (SHAP value 0.23) and "on-time check-in rate in the morning" (SHAP value 0.18). After October 25, Student A's frequency of dining in the cafeteria decreased from once a day to 0.3 times, and the on-time check-in rate in the morning decreased from 95% to 60%. The key indicators for Student B (a freshman majoring in Nursing) were "participation in online group discussions" (SHAP value 0.21) and "number of times sports venues were used" (SHAP value 0.16). In late October, participation in online group discussions decreased from 80% to 30%, and the number of times sports venues were used decreased from 3 times a week to 0 times. The system generated dynamic psychological health profiles for the two students, with a profile update cycle of 12 hours.

[0055] The real-time monitoring unit sets thresholds for abnormal fluctuations in key indicators: a week-on-week decrease of ≥40% in consumption frequency, a week-on-week decrease of ≥30% in attendance punctuality, a week-on-week decrease of ≥40% in online participation, and a week-on-week decrease of ≥80% in sports venue usage are considered abnormal. Through sliding window monitoring (window size 7 days), on October 31st, it was detected that Student A's cafeteria dinner consumption frequency decreased by 50% for a consecutive week (from an average of 1 time per day to 0.5 times), and that Student A was late for morning attendance 3 times (attendance rate decreased from 95% to 60%), triggering the risk assessment process.

[0056] The multi-dimensional risk assessment unit utilizes a hierarchical fusion assessment model. The first layer calculates the risk score for each dimension using a weighted fusion formula. The weights are determined by a team of 5 psychological experts and 10 counselors: academic stress weight 0.3, social isolation weight 0.25, behavioral abnormality weight 0.2, emotional state weight 0.15, and adaptability weight 0.1. This results in a student A's academic stress risk score of 0.3 (average midterm score of 82, no failed courses), social isolation risk score of 0.6 (0 hours of club participation, no campus activity registration record), behavioral abnormality risk score of 0.7 (abnormal spending and attendance), emotional state risk score of 0.5, and adaptability risk score of 0.4, with a comprehensive weighted score of 0.48. The second layer uses the LightGBM fusion device to combine cross-dimensional synergistic effect signals (such as the synergistic effect coefficient of behavioral abnormality and social isolation of 0.3), outputting a comprehensive risk probability score of 0.75.

[0057] The Natural Language Processing (NLP) recognition unit uses a fine-tuned MentalBERT model to analyze student A's WeChat Moments text (posted on October 26th: "I feel so sad every day, I don't want to eat or talk," and on October 30th: "I feel like I can't fit into this environment"). The high-risk content recognition layer outputs a risk level of "medium-high," the psychological and emotional state analysis layer identifies "depressive mood" (emotional intensity score 0.82), and the context and risk topic extraction layer locates the risk trigger as "poor adaptability of freshmen." Finally, a structured NLP risk report is output.

[0058] The rule engine of the early warning trigger unit presets the following trigger conditions: Red Alert (overall risk score ≥ 0.9 or NLP contains high-risk signals such as suicide or self-harm), Orange Alert (0.7 ≤ overall risk score < 0.9 and NLP contains emotional signals such as depression or anxiety), and Yellow Alert (0.5 ≤ overall risk score < 0.7 or key indicators are abnormal but there are no obvious emotional signals). Based on Student A's overall risk score of 0.75 and the presence of depressive emotional signals in the NLP, the Bayesian network calculated the early warning confidence level to be 0.92, ultimately triggering the orange alert. The system immediately pushes an early warning notification to Student A's counselor through three channels: WeChat, SMS, and the campus office system, synchronizing core risk information (overall risk score 0.75, key abnormal indicators, and NLP risk topics), generating a work order to be done within 24 hours, requiring the counselor to complete the initial intervention contact and enter the information into the system.

[0059] The classification and grading model unit adopts a multi-task hierarchical classification model. The input features are the 28-dimensional feature vector of student A, the risk scores of each dimension, the comprehensive risk score, and the NLP risk label. The model determines the severity level through ordered logistic regression. Based on the comprehensive risk score of 0.75 and two abnormal key indicators, it is judged to be at the "mild distress" level (confidence 0.89). The problem type is identified through a deep neural network for multi-label classification. Combined with information such as "social isolation risk score of 0.6" and "risk trigger is poor freshman adaptation", it is judged to be the "social withdrawal type + poor freshman adaptation type" problem type (confidence 0.85).

[0060] The intelligent matching unit first uses rule-based graph queries to initially screen out three candidate solutions from the differentiated intervention solution library that match "mild distress + social withdrawal + poor adjustment of freshmen": Solution 1 (counselor counseling + freshman adjustment group counseling), Solution 2 (AI intelligent agent social skills guidance + participation in class icebreaker activities), and Solution 3 (mindfulness meditation training + dormitory peer support). Then, it calculates the cosine similarity of the dynamic profile features of student A with historical case students, and selects 15 historical cases with a similarity ≥ 0.8. Among them, the intervention effectiveness rate of Solution 1 in 12 cases is 85%, the effectiveness rate of Solution 2 is 78%, and the effectiveness rate of Solution 3 is 65%. Finally, Solution 1 is matched as the optimal personalized intervention solution, while Solution 2 is pushed to the counselor's workbench as a backup solution.

[0061] The AI ​​diagnostic expert core of the mental health intelligent agent service module first compresses student A's dynamic profile, risk assessment results, classification and grading labels, and historical interaction records (no previous consultation records) into a 1024-dimensional structured "situational prompt" using a dedicated lightweight Transformer encoder. This prompt is then input into a finely tuned ChatGLM-6B large language model. Subsequently, based on the theme of "freshman maladjustment + social withdrawal," a vector retrieval is performed to obtain 3 authoritative knowledge points (psychological characteristics of freshman adjustment period, gentle intervention methods for social withdrawal, and key points of group counseling) and 2 similar cases from the mental health knowledge graph (containing 5000+ psychological knowledge points and 300+ typical cases). Through "thinking chain" reasoning, a diagnostic message is generated: "Hello, based on your recent performance, you may be experiencing social maladjustment during the freshman adjustment period, with a mild tendency towards social withdrawal. This is a relatively common situation among freshmen, so there is no need to be overly anxious." Intervention suggestions are also generated: "1. Have a one-on-one talk with your counselor to sort out the confusion in the adjustment process; 2. Participate in freshman adjustment group counseling to get to know classmates in the same major; 3. Try to proactively engage in simple conversations with your roommates, such as having meals or taking walks together; after verification by the fact checker that the diagnostic content and suggestions have no professional deviation, push them to the multimodal interaction unit.

[0062] The multimodal interaction unit sent a text message to student A's WeChat account while simultaneously opening a voice interaction channel. Student A replied with the text "I'm afraid to talk to strangers." The system identified "anxiety" through fine-grained sentiment analysis, and the dialogue strategy engine executed a Socratic question: "What do you worry about most when talking to strangers?" to guide the student to express further. The human-technology collaborative task distribution unit assigned a counselor to have a heart-to-heart talk with student A at 10:00 AM on November 1st, simultaneously pushing intervention plan details and the student's emotional feedback. The AI ​​agent was also arranged to push one adaptive tip to student A every day after the heart-to-heart talk, reminding him to participate in the freshman adaptive group counseling on November 3rd.

[0063] On November 1st at 10:00 AM, the counselor had a 40-minute one-on-one talk with student A, recording the key points: Student A, due to an introverted personality, was unable to quickly integrate into the group after entering university, felt lonely, and worried about being ostracized by classmates; after the talk, student A's emotions were slightly relieved, and they agreed to participate in group counseling. From November 1st to November 2nd, the AI ​​agent sent student A daily adaptation tips. On November 3rd, student A participated in freshman adaptation group counseling, and the system recorded their participation time (90 minutes in total) and number of interactions (3 group presentations). Intervention effect evaluation data showed that from November 2nd to November 8th, student A's cafeteria dinner consumption frequency recovered to an average of 0.9 times per day, their morning check-in punctuality rate increased to 88%, their comprehensive risk score decreased to 0.42, and NLP analysis of their WeChat Moments text "Met two new classmates today, felt pretty good," identified "positive emotions." These intervention process data (records of conversations, participation in group counseling), feedback information (students' emotional expressions, willingness to participate), and effect evaluation data are all fed back to the data collection unit for incremental training of the risk assessment model and classification and grading model. The weight of the "social adaptation" dimension in the model is adjusted from 0.25 to 0.28, and the threshold for abnormal fluctuations of key indicators is optimized to make the system more accurate in identifying risks during the adaptation period of new students.

[0064] The method steps in this embodiment correspond one-to-one with the above-described system workflow. Through specific technical parameter settings, data processing standards, intervention execution details, and effect verification, the actual operation process of the system is fully presented. From the implementation results, the system successfully identified student A's mild psychological distress and maladjustment issues as a freshman. Through the implementation of a personalized intervention plan, the student's abnormal behavioral indicators significantly improved, and their psychological state tended to stabilize. This verifies that the present invention can accurately capture changes in students' psychological states, provide timely warnings, and match effective personalized intervention plans, significantly improving the targeting and effectiveness of university mental health services. Simultaneously, the system optimizes model parameters through feedback data, further improving the accuracy of subsequent risk identification and intervention, forming a complete service loop.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents, characterized in that: include: The multi-source dynamic mental health profile building module is used to collect multi-source heterogeneous data, process and fuse it, and generate a dynamic mental health profile that is updated in real time. The multi-source dynamic mental health profile construction module includes a data acquisition unit, a data processing and fusion unit, and a dynamic profile generation unit. The dynamic early warning module for mental health issues is connected to the multi-source dynamic mental health profile construction module. It is used to monitor changes in key indicators of the profile in real time and combine text data analysis to achieve multi-dimensional risk assessment and graded early warning. The dynamic early warning module for mental health issues includes a real-time monitoring unit, a multi-dimensional risk assessment unit, a natural language processing and recognition unit, and an early warning triggering unit. The classification and grading intervention strategy matching module is connected to the dynamic early warning module for psychological problems. It is used to classify and grade the psychological problems identified in the early warning and match them with personalized intervention plans. The classification and grading intervention strategy matching module includes a classification and grading model unit, a differentiated intervention plan library, and an intelligent matching unit. The mental health intelligent agent service module is connected to the classification and grading intervention strategy matching module, and is used to execute personalized intervention plans and realize human-technology collaborative task distribution; the mental health intelligent agent service module includes an AI diagnostic expert core, a multimodal interaction unit, and a human-technology collaborative task distribution unit.

2. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The data acquisition unit is used to collect various multi-dimensional dynamic data from university, community, and enterprise systems.

3. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The data processing and fusion unit uses feature engineering and data mining, including surveys, averaging, and K-nearest neighbor imputation, to clean and denoise the collected data; it uses Z-score technology to standardize data of different dimensions and establishes a unified data association model based on the identification of object information.

4. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The dynamic profile generation unit uses a temporal graph neural network or dynamic knowledge graph embedding to construct a multi-dimensional dynamic association graph model. It combines a lightweight gradient boosting tree surrogate prediction model with the SHAP interpretability engine to mine a set of personalized key indicators that affect an individual's mental health and generate a dynamic mental health profile that is updated in real time.

5. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The multi-dimensional risk assessment unit adopts a hierarchical fusion assessment model. The first layer calculates independent risk scores for preset dimensions such as academic pressure and social isolation through a lightweight sub-neural network or a weighted fusion formula. The second layer uses a gradient boosting decision tree or a deep residual network fusion device to combine the risk factor scores of each dimension and the cross-dimensional synergistic effect signal to output a comprehensive risk probability score and risk attribution summary in the 0-1 range. The natural language processing recognition unit employs one or more combinations of MentalBERT, PsyBERT, or customized RoBERTa architecture models finely tuned on a mental health corpus. Through a three-layer recognition pipeline consisting of a high-risk content recognition layer, a psychological emotion state analysis layer, and a context and risk topic extraction layer, it outputs a structured NLP risk report containing high-risk signal tags, dominant emotions, and risk topics.

6. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The classification and grading model unit adopts a multi-task hierarchical classification model. First, it integrates the key indicator set of dynamic profile, risk assessment results, NLP risk labels and historical sequence information to carry out feature engineering. Then, it determines the severity level through ordered logistic regression or neural network model, identifies the problem type through deep neural network or tree model of multi-label classification, and outputs structured labels containing severity level, main type and confidence level. The differentiated intervention program library uses knowledge graph representation and causal inference analysis technology to deconstruct intervention programs into standardized and combinable operation nodes. It combines historical intervention case library to optimize the applicability conditions and recommendation weights of the programs through causal inference analysis matching propensity scores.

7. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The intelligent matching unit adopts a combination of rule-based initial screening, collaborative filtering, and similarity calculation and ranking. First, it uses rule-based graph query to screen candidate intervention schemes, then calculates the cosine similarity or Euclidean distance of the dynamic profile features of the current case and historical cases, and determines the optimal personalized intervention scheme by combining similar scheme effect data.

8. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The core of the AI ​​diagnostic expert adopts a "retrieval-enhanced generation + intelligent agent" architecture, integrating the capabilities of a large language model. First, a dedicated lightweight Transformer encoder compresses dynamic profiles, risk assessment results, classification and grading tags, and dialogue history into structured "contextual prompts." Then, based on the question, it performs vector retrieval, graph relationship retrieval, and case similarity retrieval to obtain authoritative knowledge support. Finally, it generates professional diagnostic content and intervention suggestions that conform to the template through "thinking chain" reasoning, and is verified by a fact checker.

9. The psychological risk early warning and intervention system based on multi-source heterogeneous profiling and intelligent agents according to claim 1, characterized in that: The multimodal interaction unit covers voice, text, and visual channels, and has functions such as emotion recognition, dialogue state tracking, proactive care, Socratic questioning, crisis transfer, and virtual social scenario training. It generates "real-time interactive summaries" and feeds them back to the AI ​​diagnostic expert core and dynamic profile generation unit.

10. A psychological risk early warning and intervention method based on the system described in any one of claims 1-9, characterized in that, Includes the following steps: S1. Multi-source data acquisition and preprocessing: The data acquisition unit automatically and continuously acquires multi-dimensional user data from multiple heterogeneous data sources. It uses feature engineering and data mining techniques to clean, denoise, and standardize the data. Unstructured text data is directly sent to the natural language processing and recognition unit. S2. Dynamic Mental Health Profile Construction: The data processing and fusion unit integrates the cleaned data, and the dynamic profile generation unit uses time-series graph neural network or dynamic knowledge graph embedding technology to construct a dynamic association graph model. Combined with the surrogate prediction model and interpretability engine, it mines personalized key indicator sets to generate a dynamic mental health profile that is updated in real time. S3. Psychological Risk Identification and Early Warning: The real-time monitoring unit continuously monitors key indicators of the profile, the multi-dimensional risk assessment unit calculates the comprehensive risk score, the natural language processing and recognition unit analyzes the text data and outputs a structured NLP risk report, and the early warning triggering unit generates a graded early warning signal based on the comprehensive analysis results. S4. Intervention Program Matching: The classification and grading model unit qualitatively classifies and grades the psychological problems in the warning, and the intelligent matching unit retrieves and matches the optimal personalized intervention program from the differentiated intervention program library. S5. Intervention Execution: The AI ​​diagnostic expert core of the mental health intelligent agent service module generates professional intervention content, and executes online intervention through multimodal interaction unit to interact with the object. The human-technology collaborative task distribution unit allocates tasks to the corresponding roles according to risk level. S6. Feedback and Optimization: The intervention process, feedback information, and effect evaluation data are fed back to the data acquisition unit for incremental model training and online parameter learning, thereby achieving continuous system optimization.

Images