A multi-level risk stratification agent system for psychological support for adolescents

Abstract

The application discloses a multi-level risk stratification agent system for psychological support for teenagers, and belongs to the technical field of artificial intelligence and psychological health. The system comprises a risk stratification module, a stratified intervention module, a safety and dynamic monitoring module, and a professional referral module. The risk stratification module performs a hybrid architecture classification, and divides individuals into high-risk or low-risk through automatic clinical scale screening and risk prediction based on a machine learning BPNN model; the stratified intervention module adopts a high-risk embedded question and answer path based on rules and a large language model and a low-risk multi-round dialogue path based on retrieval enhancement generation according to the stratification results; the safety and dynamic monitoring module identifies the risk level of the dialogue in real time and triggers manual review; and the professional referral module performs hierarchical recall and constrained sorting according to individual preferences, and outputs professional psychological assistance resources, thereby effectively improving the risk sensitivity, intervention pertinence and reliability of large-scale deployment of the psychological support agent.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and mental health, and in particular to a multi-level risk-stratified intelligent agent system for psychological support of adolescents. Background Technology

[0002] In recent years, mental health issues among adolescents have become increasingly prominent, while professional psychological service resources have long been in short supply, making it difficult to meet the large-scale, personalized psychological support needs. With the development of artificial intelligence technology, using dialogue agents to provide psychological support for adolescents has gradually become a research hotspot. However, existing psychological support dialogue systems still face the following key challenges in practical applications: First, insufficient risk identification and stratification capabilities. Most systems rely solely on single scales or simple rules for screening, failing to integrate automated clinical scale assessments with high-precision risk prediction based on machine learning. This results in limited sensitivity and specificity in identifying individuals with potential psychological crises, making accurate risk stratification difficult. Second, a lack of targeted intervention pathways. Existing systems typically employ a uniform dialogue strategy, failing to differentiate interventions based on individual risk levels. For high-risk individuals, there is a lack of structured crisis assessment and escalation mechanisms; for low-risk individuals, the dialogue generation process lacks effective constraints from domain knowledge, making it difficult to guarantee the professionalism and appropriateness of responses. Third, a lack of safety monitoring and closed-loop management. Open dialogues based on large language models carry the risk of generating inappropriate content or missing crisis signals. Existing systems often lack a closed-loop security mechanism covering the entire dialogue cycle, including real-time risk monitoring, manual review, and professional referral. This results in an inability to effectively and promptly protect individual safety in high-risk situations. In summary, current technologies do not yet provide a hybrid intelligent agent system that integrates automated clinical scale screening, machine learning risk prediction, tiered intervention, dynamic risk monitoring, and professional referral. Consequently, it is difficult to provide personalized and scalable psychological support to adolescents experiencing psychological distress while ensuring safety and controllability. Summary of the Invention

[0003] This invention overcomes the shortcomings of existing technologies and provides a multi-level risk-stratified intelligent agent system for psychological support of adolescents.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention discloses a multi-level risk-stratified intelligent agent system for adolescent psychological support, comprising: The system includes a risk stratification module, a stratified intervention module, a safety and dynamic monitoring module, and a professional referral module. The risk stratification module is used to perform hybrid architecture stratification, including: An automated clinical scale screening unit is used to receive the BSI-53 scale completed by an individual and calculate the total score. When the total score is greater than or equal to a preset clinical threshold, the professional referral module is triggered. When the total score is less than the preset clinical threshold, the machine learning-based risk prediction unit is triggered. The risk prediction unit is used to extract six feature variables from a preset scale filled in by an individual, input them into a pre-trained BPNN model after standardization, output risk probability values ​​and compare them with a preset classification threshold to classify individuals as high-risk or low-risk. The hierarchical intervention module includes a high-risk embedded question-and-answer unit and a low-risk retrieval-based multi-turn dialogue unit. The high-risk embedded question-answering unit is used to conduct natural language assessments on high-risk individuals using a rule-based predefined dialogue framework, focusing on suicide risk, self-harm behavior, and somatization symptoms. It then calls a large language model to generate supportive responses and escalation judgments, and triggers manual review or redirection to the low-risk retrieval-based multi-turn dialogue unit based on the judgment results. The low-risk retrieval-based multi-turn dialogue unit is used to conduct multi-turn dialogues with low-risk individuals using a retrieval-enhanced generation method. Based on the individual's current intervention stage and input content, it retrieves similar historical dialogue records from a knowledge base that includes peer mentor dialogues and public psychological counseling dialogues, and generates personalized responses. The security and dynamic monitoring module is used to acquire multi-turn dialogue context in real time, call the large language model to perform risk classification through preset risk judgment prompt words, and trigger manual review when any risk level is identified. The professional referral module is used to collect individual preferences through multi-round dialogue when a referral trigger signal is received, and to perform hierarchical recall and constraint sorting in the referral resource database to output recommended professional psychological assistance resources.

[0005] Preferably, the preset clinical threshold is 63 points.

[0006] Preferably, in the risk prediction unit, the six feature variables are derived from subdomain scores of the Life Events Self-Rating Scale, the Perceived Social Support Scale, and the Emotional Intelligence Scale; the BPNN model is a single-hidden-layer feedforward neural network with six features in the input layer, three neurons in the hidden layer using the ReLU activation function, and the output layer using the Sigmoid function to output the risk probability; the standardization is based on the parameters of the training dataset; and the preset classification threshold is 0.5.

[0007] Preferably, the specific rules for sequential evaluation in the high-risk embedded question-answering unit are as follows: First, a suicide risk screening is conducted. If an individual expresses suicidal ideation or behavior, a manual review is triggered directly. Otherwise, a self-harm behavior screening will be conducted, and if self-harm behavior is found, a manual review will be triggered. Otherwise, a somatization assessment is performed, which includes four progressive questions Q1, Q2, Q3, and Q4, where Q1 addresses somatization symptoms, and Q2, Q3, and Q4 address anxiety and depression symptoms, depression symptoms, and anxiety symptoms, respectively. When Q1 is positive and any one of Q2, Q3 or Q4 is positive, manual review is triggered; Manual review is triggered when Q1 is negative and Q2, Q3 and Q4 are all positive. When all four rounds of Q1 are negative, the individual is returned to the low-risk retrieval multi-turn dialogue unit.

[0008] Preferably, in the low-risk retrieval-based multi-turn dialogue unit, the knowledge base construction and retrieval process includes: Clean up counseling dialogue records and make public the adolescent psychological counseling dialogue dataset, remove cases containing clear mental illness diagnoses and self-harm and suicide content, and categorize them by intervention stage and theme; The dialogue text is vectorized and stored in a vector database, with individual statements as the core and adjacent sentences combined to form retrieval units. During the search, the peer mentor dialogue database is searched according to the individual's current intervention stage. If the results are insufficient, the individual's input is subject to topic identification and supplemented by searching the public dataset of the corresponding topic. The current individual input and candidate dialogue fragments are vector-encoded, and the cosine similarity score is calculated. After sorting in descending order and removing duplicates, the top 3 are selected as the reference for generating the response.

[0009] Preferably, the low-risk retrieval-based multi-turn dialogue unit further includes a behavioral task generation constraint mechanism: During the first consultation each week, at least one behavioral task is generated based on the comprehensive analysis of multiple rounds of dialogue, and the task content is written into the individual's long-term memory module. Within the same intervention period, new behavioral tasks should not be generated repeatedly. Instead, historical task information should be used for tracking, feedback, and encouragement. When the task is detected to be completed or the next intervention cycle is entered, the task generation permission is re-enabled.

[0010] Preferably, the safety and dynamic monitoring module is specifically used for: Extract the most recent k=6 rounds of interaction between the individual and the agent as a context window. Input the context and predefined risk classification criteria into the large language model for structured reasoning and output the risk level L1, L2 or L3. L1 indicates extremely negative emotions but no clear intention to self-harm or commit suicide. L2 indicates the existence of suicide or self-harm risk but has not entered the execution stage. L3 indicates the existence of immediate life danger or the implementation of self-harm or suicide behavior. When any risk level is identified, the agent's response is paused and a manual review is triggered. An expert will assess the situation based on the complete dialogue context. If the risk is manageable, the interaction will resume; otherwise, the professional referral module or emergency intervention will be triggered.

[0011] Preferably, the professional referral module includes a referral resource database, which stores manually verified psychological service hotlines and institution information; the referral trigger signal includes: the BSI-53 total score is greater than or equal to the preset clinical threshold, or the individual is determined by manual review in the safety and dynamic monitoring module to be unsuitable for further intervention; The multi-turn dialogue collects individual preferences including population type, regional information, and cost preferences; The hierarchical recall is carried out in the following order of priority: joint provincial and municipal matching, city matching, province matching, provincial record matching, and national resource matching. When the matching fails, the city field is subjected to suffix removal and normalization as a fallback match. After recalling the candidate list, the candidate list is reordered with the large language model constrained by individual preference inputs, allowing selection only from the candidate list, and outputting structured information of up to 5 matching institutions.

[0012] This invention also discloses a multi-level risk stratification method for adolescent psychological support, applicable to any of the systems described herein, comprising: Receive the completed BSI-53 scale from the individual, calculate the total score, and refer the individual to professional services if the total score is greater than or equal to the preset clinical threshold; otherwise, proceed to the next step. Six feature variables are extracted from the pre-defined scale filled in by the individual, standardized and input into the pre-trained BPNN model to obtain the risk probability value and compare it with the threshold. If it is lower than the threshold, the low-risk retrieval multi-turn dialogue unit is used for intervention; otherwise, the high-risk embedded question-answering unit is used for intervention. During the intervention process using the aforementioned low-risk retrieval-based multi-turn dialogue unit, personalized responses are generated by combining the knowledge base and historical dialogues, and dialogue risks are monitored in real time. During the intervention using the high-risk embedded question-answering unit, suicide risk, self-harm behavior and somatization symptoms are assessed sequentially through natural language dialogue. Based on the large language model, it is determined whether to escalate and trigger manual review. Throughout the intervention process, the context of multi-turn dialogues is classified in real time for risk assessment. When a risk is identified, manual review is triggered, and referral or resumption of interaction is performed based on the review results. When the referral criteria are met, individual preferences are collected and hierarchical recall and constrained ranking are performed in the referral resource database to output recommended professional institutions.

[0013] Preferably, during the intervention process of the high-risk embedded question-and-answer unit, the somatization symptom assessment adopts four progressive questions. If the first question is positive, any subsequent question being positive will trigger manual review. If the first question is negative, all subsequent questions must be positive to trigger manual review. If all are negative, the process returns to the low-risk retrieval-based multi-turn dialogue unit.

[0014] This invention addresses the technical deficiencies in the prior art and offers the following advantages: By employing a hybrid architecture combining automated clinical scale screening and machine learning prediction for risk stratification, coupled with rule-based high-risk structured assessment and enhanced generative low-risk dialogue, supplemented by full-time semantic risk monitoring and manual review, and preference-driven constrained referral recommendations, this invention achieves accurate identification, differentiated intervention, and safe closed-loop management of adolescents with psychological distress. While ensuring safe and controllable human-computer interaction, the system enhances the risk sensitivity, intervention targeting, and professional constraint capabilities of large language models in mental health services, effectively reducing the risk of high-risk missed diagnoses and misdiagnoses, and enhancing the scalability and deployment reliability of large-scale psychological support services. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0016] Figure 1 A diagram illustrating the architecture of a multi-level risk-stratified intelligent agent system for adolescent psychological support; Figure 2 A flowchart for a multi-level risk stratification approach to psychological support for adolescents. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention.

[0018] like Figure 1 , Figure 2 As shown, this invention provides a multi-level risk stratification intelligent agent system and method for adolescent psychological support. The system adopts a hybrid architecture, organically integrating automated clinical scale screening, machine learning-based risk prediction, structured assessment and retrieval-enhanced generative dialogue based on large language models, dynamic safety monitoring, and professional referral, achieving intelligent psychological support throughout the entire process from initial risk screening and stratified intervention to closed-loop safety management.

[0019] System Overall Architecture In one embodiment, the intelligent agent system of the present invention is implemented using a microservice architecture, including a front-end interaction layer, a back-end business logic layer, a model service layer, and a data storage layer. The front-end is built on the React framework and is responsible for presenting questionnaires, dialogue interfaces, and the display of structured information. The back-end is built on Python's FastAPI framework and is responsible for business process orchestration, API calls, and data processing. The large language model uses the Qwen model, which provides language understanding and generation capabilities for multiple stages such as high-risk assessment, low-risk dialogue generation, security monitoring, and referral optimization through a unified API interface. The vector database uses Chroma to store and retrieve psychological counseling dialogue fragments. All functional modules communicate through standardized interfaces to ensure high cohesion and low coupling within the system.

[0020] Risk stratification module The risk stratification module is the entry point for this system to perform hybrid architecture stratification, including an automated clinical scale screening unit and a machine learning-based risk prediction unit.

[0021] The automated clinical scale screening unit receives completed BSI-53 questionnaires from individuals and calculates the total score. BSI-53 (Brief Symptom Inventory-53) is a widely used brief symptom scale in psychology, psychiatry, and nursing to assess an individual's psychological symptoms over the past 7 days. The scale contains 53 items covering nine major symptom dimensions: somatization, obsessive-compulsive disorder, interpersonal sensitivity, depression, anxiety, hostility, phobias, paranoia, and psychosis. Each item uses a 5-point Likert scale from 0 (completely not) to 4 (extremely). The front-end renders the questionnaire using React and Formic components. After registration, the system automatically pushes the BSI-53 questionnaire to the client. After submission, the front-end sends the raw scores of the 53 items in JSON format to the back-end interface. The back-end receives the data via FastAPI and sums the scores of each item according to predefined scoring rules to obtain the total BSI-53 score.

[0022] The system has a preset clinical threshold, with a preferred value of 63. When the BSI-53 total score is greater than or equal to 63, it indicates that the individual's psychological symptoms are relatively severe, exceeding the safety boundary of automated intervention. At this time, the automated clinical scale screening unit triggers the professional referral module and generates a manual follow-up task, which is then carried out by professionals to refer the individual to offline professional services. When the BSI-53 total score is less than 63, the individual enters the next step of the risk prediction process.

[0023] The machine learning risk prediction unit is used to perform more accurate risk classification for individuals who pass the initial BSI-53 screening. This unit is implemented based on a pre-trained BPNN (Backpropagation Neural Network) model. The model structure is a single-hidden-layer feedforward neural network. The input layer contains 6 feature variables, the hidden layer has 3 neurons using the ReLU activation function, and the output layer has a single neuron using the Sigmoid activation function, outputting a risk probability value between 0 and 1. The model was trained on 3038 samples using the Adam optimizer and the binary cross-entropy loss function, and underwent internal validation (C-index 0.9120) and external validation (539 samples, C-index 0.9749), demonstrating excellent discriminative performance and generalization ability.

[0024] The model's six input feature variables are derived from subdomain scores of three pre-defined scales: the three subdomain scores (SLE.D1, SLE.D4, and SLE.D5) of the Self-Rating Life Events Scale (ASLEC); the one subdomain score (SS.D3) of the Perceived Social Support Scale (PSSS); and the two subdomain scores (EI.D2 and EI.D3) of the WLEIS (Willing Emotional Intelligence Scale). These six features were selected through SHAP feature importance analysis combining random forest and XGBoost models.

[0025] During the inference phase, the system extracts relevant item scores from the baseline questionnaire completed by the individual and constructs a 6-dimensional feature vector according to predefined mapping rules. To ensure consistency with the training process, all feature vectors are normalized using the Z-score normalization method based on the mean and standard deviation of the training dataset. The normalized feature vectors are input into the trained BPNN model, and a risk probability value is calculated through forward propagation. The system compares this probability value with a preset classification threshold (preferably 0.5): if the probability value is below 0.5, the individual is classified as low-risk; if the probability value is greater than or equal to 0.5, the individual is classified as high-risk.

[0026] To achieve seamless integration of the predictive model and the agent, this embodiment encapsulates the BPNN model as a standalone web-based risk calculator and embeds it into the system workflow via a standardized API interface. The calculator uses ShinyforPython to build the interactive interface and is deployed in a cloud environment using Docker containerization. The system automatically invokes the calculator after an individual completes a baseline questionnaire to complete risk stratification decisions. The model operates using offline training and online inference. During the inference phase, only forward computation is performed without parameter updates. Furthermore, the calculator shares the same model parameters, feature processing procedures, and inference logic with the main system workflow, ensuring consistency and reproducibility of results across different environments.

[0027] Stratified intervention module The tiered intervention module is the core intervention execution part of the system, which includes a high-risk embedded question-and-answer unit and a low-risk retrieval-based multi-turn dialogue unit. Based on the risk tiering results, individuals are guided to different intervention paths.

[0028] The high-risk embedded question-and-answer unit targets individuals classified as high-risk. It employs a rule-based, predefined dialogue framework, sequentially assessing three dimensions—suicide risk, self-harm behavior, and somatization symptoms—using natural language. This assessment strategy integrates key risk factors from the field of mental health crisis identification: suicide is directly related to life safety and is given the highest priority; self-harm behavior is a behavioral manifestation of mood regulation disorders and is closely related to suicide; somatization symptoms reflect implicit expressions of psychological distress. The standardized screening scale questions for these three dimensions are transformed into a natural language dialogue format, compressed into single-round open-ended questions to reduce individual defensiveness and increase the acceptability of expression. All questions are reviewed by mental health professionals.

[0029] The specific assessment rules are as follows: First, a suicide risk screening is conducted. A large language model is used to semantically understand the individual's responses. If the individual is determined to have suicidal ideation or behavior, manual review is triggered directly. If the suicide screening is negative, the self-harm behavior screening proceeds; if self-harm behavior is found, manual review is triggered. If the self-harm screening is also negative, the somatization assessment proceeds. The somatization assessment includes four progressive questions, Q1, Q2, Q3, and Q4, targeting somatization symptoms, anxiety and depression symptoms, depression symptoms, and anxiety symptoms, respectively. When Q1 is positive, manual review is triggered if any of Q2, Q3, or Q4 is positive; when Q1 is negative, manual review is triggered only if all of Q2, Q3, and Q4 are positive; if all of Q1 to Q4 are negative, it indicates that the individual does not currently exhibit significant somatization-related psychological risk, and the individual is returned to the low-risk retrieval-based multi-turn dialogue unit.

[0030] In this unit, the system uses a unified prompt template, inputting individual responses and necessary dialogue context into a large language model. The model simultaneously generates supportive responses and structured judgment results. Supportive responses are used to maintain empathy and fluency in the dialogue; judgment results are output in a predefined JSON format, containing assessment conclusions for each dimension and a flag indicating whether manual review has been triggered. An example prompt is as follows: "You are a professional adolescent mental health support provider. Based on the following dialogue, please provide a brief empathetic response and professional assessment of the client in the three dimensions of suicide risk, self-harm behavior, and somatization symptoms, and output the assessment results in JSON format, including four fields: 'suicide_risk' (Boolean), 'self_harm' (Boolean), 'somatization' (Boolean), and 'need_manual_review' (Boolean)." The low-risk retrieval-based multi-turn dialogue unit targets low-risk individuals and those returning from high-risk pathways, employing a retrieval-enhanced generation (RAG) approach to provide multi-turn dialogue support. The core of this unit is a proprietary knowledge base focused on adolescent mental health support. The knowledge base was constructed as follows: the text data came from two sources: anonymized human peer mentor intervention project consultation dialogues (approximately 5187 rounds) and publicly available online adolescent mental health counseling dialogue datasets (approximately 44396 rounds), covering topics such as family, interpersonal relationships, self-growth, and emotion regulation. Public datasets included the open-source Lingxin large model fine-tuning dataset and the digital twin dialogue dataset for mental health counselors released by institutions such as the Future Technology Institute of South China University of Technology. During construction, the data underwent rigorous cleaning: topics related to marriage and relationships and therapy were excluded; cases with clear diagnoses of mental illnesses such as depression and anxiety were removed; and cases expressing extreme despair, suicidal tendencies, or self-harm tendencies were deleted. The cleaned corpus is centered on individual statements, combining adjacent sentences to form retrieval units, essentially piecing together a single user's statement with its context to create an independent segment. Each retrieval unit is vector-encoded using a text embedding model (such as text2vec-large-chinese) to generate a dense vector of fixed dimensions, which is then stored in the Chroma vector database. The vector database is logically partitioned according to data source and attributes: the peer mentor dialogue database is divided into three sub-databases according to intervention stages: T0 (initial acquaintance and relationship building), T1-T3 (intermediate intervention), and T4 (termination and consolidation); the public dataset is divided into sub-databases according to four themes: family, interpersonal relationships, self-growth, and emotion regulation.

[0031] When an individual initiates a consultation, the system first obtains their current intervention stage and performs a similarity search in the corresponding stage sub-database of the peer mentor dialogue database. If the number of search results is less than a preset threshold (e.g., 2), a supplementary search mechanism is triggered: a keyword and rule-based topic classifier is invoked to identify the topic of the individual's current input, determining whether it belongs to a category such as family, interpersonal relationships, self-growth, or emotion regulation, and then searching in the corresponding topic's public data sub-database. The search process encodes the current individual input and candidate dialogue fragments in the database into vector representations e. q and e di Calculate the cosine similarity score: , The system sorts the records in descending order based on similarity scores, and after deduplication, selects the top 3 most relevant historical dialogue records as a reference for generating responses.

[0032] When generating personalized responses, the system inputs retrieved historical dialogue fragments, the current dialogue context, and predefined personality and role cue words into the Big Language Model. An example of a role cue word is: "You are a warm, inclusive, and psychologically knowledgeable adolescent peer supporter. Please refer to the following dialogue fragments in similar situations and, based on the visitor's current expression, provide an empathetic, empowering, and supportive response, avoiding direct diagnostic or treatment suggestions." The Big Language Model then generates personalized responses that conform to professional standards and supportive goals.

[0033] The low-risk retrieval-based multi-turn dialogue unit also incorporates a behavioral task generation constraint mechanism to achieve goal-oriented continuous intervention. During the initial consultation each week, the system comprehensively analyzes all dialogue content from that week, extracts key issues and actionable behaviors, and automatically generates at least one specific behavioral task, such as "This week, please try to record three small things that make you happy or grateful each day." After task generation, the system structures it into a JSON object, including task description, generation time, and completion status, and writes it into the individual's long-term memory module. Within the same intervention cycle (usually a week), the system prohibits the generation of new behavioral tasks. When the individual initiates a dialogue again, the system automatically retrieves historical task information from long-term memory and tracks, inquires about, provides feedback on, and encourages the task execution in natural language, such as "Last time we agreed to try recording three small things that made you happy; how has it progressed these past few days?" When the system detects through dialogue analysis or the individual's active confirmation that the task has been completed, or when the timestamp enters the next intervention cycle, the task generation permission is reopened.

[0034] Safety and dynamic monitoring module The security and dynamic monitoring module operates throughout the entire dialogue process, serving as a crucial barrier to ensure safe and controllable human-computer interaction. This module acquires multi-turn dialogue context in real time and uses preset risk assessment prompts to invoke a large language model for semantic risk classification.

[0035] In its implementation, the system maintains a sliding window, extracting the interaction content between the individual and the agent in the most recent k=6 rounds as the context window, which, along with a predefined risk grading standard, is input into the large language model. The risk grading standard divides the severity of the crisis into three levels: L1 indicates extremely negative emotions but no clear suicidal or self-harming intentions; L2 indicates a risk of suicide or self-harm but not yet at the execution stage; and L3 indicates an immediate danger to life or ongoing or imminent self-harm or suicide. The judgment criteria particularly emphasize understanding the current context, requiring the model to distinguish between past experiences and the current state, and between hypothetical statements and actual intentions, thereby reducing misjudgments of non-crisis statements. The large language model is guided to output structured results, including a risk level field and a summary of the judgment criteria.

[0036] When the system identifies any risk level (L1, L2, or L3), it pauses the current dialogue response and pushes a review task containing the full dialogue text and risk assessment results to the manual review queue. Manual review is performed by experts with professional psychological counseling qualifications and who have undergone standardized system training. Review experts view the complete dialogue context through the management backend and conduct a comprehensive assessment based on their professional judgment. If the expert determines the risk is manageable, the dialogue restriction is lifted through the backend, the system resumes human-computer interaction, and continues to generate supportive responses; if a persistent or escalating risk is identified, the system triggers the professional referral module or emergency intervention process. The emergency intervention process includes sending alert notifications to pre-set emergency contacts and prompting individuals to call the psychological crisis intervention hotline.

[0037] Professional referral module The professional referral module serves as a safety net for the system, directing individuals to specialized mental health resources when they are no longer suitable for agency-based psychological intervention. This module maintains a rigorously manually verified database of referral resources, containing information on approximately 500 mental health service hotlines and service providers. Verification includes manually dialing hotline numbers and verifying each link to an organization's website to ensure data validity and reliability.

[0038] Referral trigger signals come from two sources: first, the automated clinical scale screening unit determines that the BSI-53 total score is greater than or equal to the preset clinical threshold (63 points); second, the safety and dynamic monitoring module manually reviews and determines that the individual is not suitable to continue the current intervention process.

[0039] Once the referral process is triggered, the system first sends standardized care and referral instructions to the individual. Then, through multiple rounds of dialogue, it sequentially collects three aspects of preference information: population type (teenager, student, or others), regional information (the individual can describe their region using free text), and cost preference (free or paid services). For the free text regional information, the system uses a large language model for structured parsing, mapping it to standardized province abbreviations and city full names. For example, inputting "I am in Guangzhou" can be parsed as province "Guangdong" and city "Guangzhou"; when incomplete information is input, the system completes it through follow-up questions.

[0040] After completing preference collection, the system performs hierarchical regional recall in the referral resource database, prioritizing the following order from highest to lowest: provincial and municipal joint exact matching, city name exact matching, province matching, provincial record matching (i.e., only province matching with the city field empty or a provincial record), and nationwide resource matching. When none of the above hierarchical matching yields enough candidates, the system performs suffix removal normalization on the city field in the database (e.g., removing suffixes such as "city" and "district") before performing fallback matching to further improve the recall rate. The list of candidate institutions obtained from the recall, along with individual preference information, is input into the constrained large language model fine ranking module. This module employs strict instruction constraints, requiring the large language model to sort and filter only from the provided candidate list based on matching degree and preference fit, prohibiting the generation of any institution names outside the list or rewriting of structured fields, thereby ensuring the controllability of the recommendation results. Ultimately, the system outputs up to 5 of the most suitable professional institutions, presented to the individual in the form of structured fields, including: province, city, hospital or institution name, hotline name, target audience, service hours, hotline number, whether it is free, and data source.

[0041] Dynamic scale assessment and closed-loop intervention To further enhance the dynamic monitoring and closed-loop management of the intervention process, this system deeply embeds the BSI-53 scale assessment into the intervention cycle. In addition to the initial screening upon enrollment, in the low-risk intervention pathway, the system automatically pushes the BSI-53 reassessment questionnaire at the end of each weekly dialogue. After individual submission, the system recalculates the total score: if the BSI-53 total score is greater than or equal to 63, the professional referral module is automatically triggered; if the total score is still less than 63, the individual continues to participate in the next round of dialogue intervention. This mechanism ensures that even if the initial risk is low, if psychological symptoms worsen during the intervention process, they can be promptly identified and referred to professional services, forming a complete safety loop.

[0042] Implementation Examples of Multi-Level Risk Stratification Method Corresponding to the above system, this invention also provides a multi-level risk stratification method for psychological support of adolescents. The method flow is as follows: Step S1: Receive the completed BSI-53 scale from the individual, calculate the total score. If the total score is greater than or equal to the preset clinical threshold (e.g., 63 points), refer the individual to professional services; if the total score is less than the preset clinical threshold, proceed to step S2.

[0043] Step S2: Extract the individual's scores from six specified subdomains of the Life Events Self-Rating Scale, Perceived Social Support Scale, and Emotional Intelligence Scale to form a feature vector. After Z-score standardization based on the training dataset parameters, input the vector into a pre-trained BPNN model to obtain a risk probability value. Compare this probability value with a preset classification threshold (0.5): if it is lower than the threshold, intervene using a low-risk retrieval-based multi-turn dialogue unit; otherwise, intervene using a high-risk embedded question-answering unit.

[0044] Step S3: During the low-risk retrieval-based multi-turn dialogue unit intervention, the system retrieves similar historical dialogue records from a knowledge base containing peer mentor dialogues and public psychological counseling dialogues, based on the individual's intervention stage and input content. Cosine similarity is calculated, and the top three records are used as references. Personalized responses are generated using a large language model. Simultaneously, at least one behavioral task is generated and written into long-term memory during the first consultation each week. Within the same intervention cycle, task progress is tracked and feedback is provided, but repeated generation is avoided. The entire process is monitored in real-time by a safety and dynamic monitoring module to assess dialogue risks.

[0045] Step S4: During the intervention using the high-risk embedded question-and-answer unit, suicide risk, self-harm behavior, and somatization symptoms are assessed sequentially via natural language dialogue. The somatization symptom assessment uses four progressive questions. If the first question is positive, any subsequent question being positive triggers manual review; if the first question is negative, all subsequent questions must be positive to trigger manual review; if all questions are negative, the process returns to the low-risk retrieval-based multi-turn dialogue unit. Positive results in other assessment stages directly trigger manual review.

[0046] Step S5: Throughout the intervention process, the safety and dynamic monitoring module performs real-time risk classification on the context of the last 6 rounds of dialogue. When L1, L2 or L3 risk levels are identified, manual review is triggered. Based on the expert review results, the interaction is restored or a professional referral is triggered.

[0047] Step S6: When the referral conditions are met, collect the individual's population type, region and cost preference, perform hierarchical recall and constrained large language model fine ranking in the referral resource database, and output information on up to 5 suitable professional institutions.

[0048] In another preferred embodiment of the invention, it further includes: When the safety and dynamic monitoring module is triggered to perform risk classification each time, the current weight vector is normalized to obtain the probability distribution vector of each expert selected in the current round; based on the probability distribution vector, the active expert for this round is determined from the expert pool by a probability ratio random selection method, the risk judgment prompt word template corresponding to the active expert for this round is called, the dialogue context is input into the large language model and the risk level judgment result for this round is obtained; The risk level assessment results of this round are compared with the actual risk labels fed back in the subsequent manual review stage to calculate the indicated loss of the activated experts in this round. The indicated loss is defined as follows: if the risk level assessment results of this round are consistent with the actual risk labels, the indicated loss is zero; if they are inconsistent, the indicated loss is a non-zero constant value. Based on the indicated loss, the unbiased loss estimation vector of the activated expert in this round is determined, wherein the estimated value of the activated expert in this round in the corresponding dimension of the unbiased loss estimation vector is the indicated loss divided by the probability of the activated expert in this round being selected in the probability distribution vector, and the estimated values ​​of the other experts in the corresponding dimension of the unbiased loss estimation vector are all set to zero. Using the unbiased loss estimation vector, according to the exponential weighted average update rule, the weight values ​​corresponding to each expert in the previous round of weight vector are multiplied by the negative unbiased loss estimation multiple of the preset learning rate factor, and the weights are updated by expert dimension to generate the updated weight vector for the next round of risk classification. During multiple rounds of interaction, the above steps are repeated continuously so that in the next round of risk classification, the expert with the highest current weight value corresponds to the expert in the expert pool with the smallest historical cumulative indication loss, so as to asymptotically minimize the cumulative regret of safety monitoring misjudgments.

[0049] It should be noted that, considering the high individual variability and context-dependent nature of adolescent psychological expression, a single, fixed risk assessment prompt template may be insufficiently sensitive to certain implicit expressions or crisis signals in specific cultural contexts, leading to misjudgments or omissions. To overcome this deficiency, this implementation first normalizes the saved current weight vector each time the safety and dynamic monitoring module is triggered to perform risk grading. Each dimension of this vector corresponds to an expert in the expert pool, and normalization yields the probability distribution vector of each expert being selected in the current round. For example, if the expert pool contains three experts and the current weight vector is (1.2, 0.8, 0.6), the normalized probability distribution vector is (0.462, 0.308, 0.230). Then, based on this probability distribution vector, the system uses a probability-proportional random selection method to determine the active experts for this round from the expert pool. That is, the probability of each expert being selected is equal to its normalized weight, while retaining a certain degree of exploratory nature to avoid prematurely falling into local optima. Next, the system calls the risk assessment prompt word template corresponding to the activated expert, inputs the recent dialogue context into the large language model, and obtains the risk level assessment result for this round.

[0050] Subsequently, the risk level assessment results of this round are compared with the actual risk labels provided in the subsequent manual review stage to calculate the indicated loss of the activated experts in this round. The indicated loss is used to quantify the deviation between the risk grading result and the actual situation. For example, it can be agreed that if the assessment result matches the actual risk label, the indicated loss is 0; if they do not match, the indicated loss is a constant value of 1 (or other preset non-zero value) to concisely reflect the effectiveness of a single decision. Based on the indicated loss, the unbiased loss estimation vector of the activated experts in this round is further determined. Its purpose is to construct an important statistic that can unbiasedly estimate the expected loss of all experts, providing a basis for subsequent weight updates, given that only the losses of the selected experts are observed. The dimension of this vector is the same as the number of experts. The estimated value for the dimension corresponding to the activated expert in this round is set to the indicated loss divided by the probability that the expert was selected in this round, while the estimated values ​​for the dimensions corresponding to the other unactivated experts are all set to zero. Taking the three experts mentioned above as an example, if the second expert is activated in this round of probabilistic selection, and the selection probability of this expert is 0.308, and their judgment result is inconsistent with the manually reviewed label, indicating a loss of 1, then the unbiased loss estimation vector for this round is (0, 1 / 0.308≈3.25, 0). This construction method ensures that, in the sense of probabilistic expectation, the expected value of this estimation vector is equal to the true indicating loss vector of each expert, thus providing the correct gradient direction for overall optimization.

[0051] After obtaining the unbiased loss estimation vector, the weight vector from the previous round is updated sequentially along the expert dimension according to the exponentially weighted average update rule. Specifically, the update method involves multiplying the weight value corresponding to each expert in the previous round's weight vector by the negative unbiased loss estimation multiple of a preset learning rate factor, thereby reducing the weight of poorly performing experts in this round and indirectly increasing the probability of selecting other experts in the future. This update process can be expressed as: New weight = Old weight × exp(-η × Unbiased loss estimation), where η is the learning rate parameter, controlling the weight update step size. After one round of updates, the updated weight vector for this round is generated and used for expert selection in the next round of risk grading.

[0052] During multiple rounds of interaction, the above steps are repeated continuously, causing the weight distribution to gradually tilt towards the expert with the lowest historical cumulative indication loss. Through this adaptive learning mechanism, the system can dynamically approximate the optimal risk assessment strategy in a data-driven manner without prior knowledge of the optimal expert, and theoretically guarantees long-term monitoring performance with an upper bound of sublinear cumulative regret. As a result, the risk classification accuracy of the safety and dynamic monitoring module continuously improves with actual operation, further reducing the false alarm rate in high-risk situations and the false alarm rate in normal interactions, thus enhancing the safety and controllability of the entire human-computer interaction system.

[0053] In another preferred embodiment of the invention, it further includes: Extract the multi-source features generated in the current interactive session and construct a set of concept nodes for a fuzzy cognitive graph. The concept nodes include at least: scale feature nodes mapped from individual life event stress scores, perceived social support scores, and emotion regulation scores obtained from the risk stratification module; dialogue semantic intensity nodes mapped from the negative emotion intensity value and self-harm semantic tendency value corresponding to the most recent dialogue round obtained from the safety and dynamic monitoring module; and monitoring risk level nodes obtained by fuzzifying the positive assessment result sequence of somatization symptoms output by the high-risk embedded question-and-answer unit. Define a trapezoidal membership function for each concept node and convert the scores of each item into the initial activation membership of the concept node; and load the fuzzy cognitive graph causal weighting matrix pre-installed in the system knowledge base; The initial activation membership degrees of each concept node are used to form an initial activation vector. The initial activation vector is synchronously iteratively updated using the causal weighting matrix. In each iteration, based on the activation membership degrees of each concept node in the previous round and the weights of each incoming edge in the causal weighting matrix, the updated activation membership degrees of each concept node in the current round are calculated using a cumulative summation with saturation limit and a nonlinear compression function. This process continues until the maximum change in the activation membership degrees of each node between the two rounds is lower than the preset convergence tolerance, thus obtaining the final activation membership degree of each concept node. In the final activation membership, concept nodes whose activation membership exceeds the preset vulnerability threshold are marked as potential vulnerable nodes; For each potentially vulnerable node, starting from that node, a depth-first path search is performed along the outgoing edges with weights greater than zero in the causal weighted matrix. The product of the activation membership increment of the concept nodes passed through each reachable path and the corresponding edge weight is calculated as the chain effect strength of the corresponding path. The output is the multiple chain risk paths triggered by the potentially vulnerable node and the chain effect strength ranking of each path. From the chain risk path with the strongest chain impact, concept nodes located in the middle of the path and with adjustable incoming edge weights in the causal weighting matrix are extracted in reverse as intermediate controllable nodes, and preventive dialogue nodes corresponding one-to-one with each intermediate controllable node are generated; each preventive dialogue node contains intervention topic text and target activation direction identifier with the goal of reducing the activation membership of the corresponding concept node. The system sequentially injects the generated sequence of preventative dialogue nodes into the context of the responses to be generated in low-risk retrieval-based multi-turn dialogue units, in order to buffer the cascading risks caused by the triggering of potentially vulnerable nodes.

[0054] It should be noted that in real-world psychological support scenarios, an individual's life stress, perceived social support, and emotional regulation abilities are not isolated from the intensity of emotions expressed in the current conversation and their tendency towards self-harm. Rather, they are interconnected and mutually causal; the deterioration of one risk factor may trigger a chain reaction along a specific path. If the system only makes isolated threshold judgments on each risk indicator, it is difficult to anticipate the trend of cascading risks and may miss the optimal intervention window. To address this, this embodiment constructs a fuzzy cognitive graph, integrating heterogeneous features generated by multiple modules and rounds into a unified causal network. Based on this network, it actively searches for potential chain risk paths, achieving an upgrade from passive monitoring to proactive prediction. In specific implementation, the multi-source features generated in the current interactive session are first extracted to construct a set of concept nodes for the fuzzy cognitive graph. The concept nodes include at least three categories: First, scale feature nodes, specifically formed by normalizing and mapping individual life event stress scores, perceived social support scores, and emotion regulation scores obtained from the risk stratification module; Second, dialogue semantic intensity nodes, specifically formed by the negative emotion intensity value and self-harm semantic tendency value corresponding to the most recent rounds of dialogue obtained from the safety and dynamic monitoring module. The former is obtained by scoring the emotional polarity of individual speech using a large language model, and the latter is obtained by jointly calculating with a preset self-harm keyword dictionary and semantic similarity; Third, monitoring risk level nodes, specifically obtained by fuzzifying the sequence of positive assessment results of somatization symptoms output by the high-risk embedded question-and-answer unit. For example, mapping the positive or negative results of each item in multiple consecutive assessments to a continuous membership value between 0 and 1 to characterize the persistence of somatization risk. It should be noted here that fuzzy cognitive graph is a soft computational model that combines fuzzy logic with cognitive graph. Its core lies in using a directed weighted graph to express the causal relationship between concept nodes and using a membership function to characterize the fuzziness of the state of each concept node. In this embodiment, the fuzzy cognitive graph causal weighted matrix pre-installed in the system knowledge base is essentially a domain knowledge-driven prior relation table. Its rows and columns correspond to each concept node, and the values ​​of the elements in the matrix represent the causal influence strength of the column node on the row node. Positive weights indicate positive promoting relationships, negative weights indicate negative inhibiting relationships, and zero values ​​indicate no direct causal relationship. This matrix is ​​initially assigned values ​​by experts in the field of psychology based on theories related to adolescent mental health, and calibrated through data fitting in a small-scale pre-experiment. It is then loaded into the system as static parameters for online inference.

[0055] After constructing the concept node set, the system defines a trapezoidal membership function for each concept node. This function is completely determined by four inflection point parameters. The setting of the upper and lower bases of the trapezoid allows for flexible characterization of the transition range of fuzzy linguistic values ​​such as "low," "medium," and "high." Based on this membership function, the score or value corresponding to each concept node is converted into the initial activation membership degree of that node, with a value range of [0,1]. The higher the value, the more significant the risk characteristic represented by the node is in the current state. The initial activation membership degrees of all nodes together form the initial activation vector.

[0056] Subsequently, the system uses a pre-set fuzzy cognitive graph causal weighting matrix to perform synchronous iterative updates on the initial activation vector. In each iteration, for any target node, the cumulative causal input of that node is obtained by multiplying the activation membership degrees of all incoming edge source nodes from the previous iteration with their corresponding causal weights and summing the products. To avoid unbounded divergence or premature saturation of activation values ​​due to multiple iterations, the system applies a saturation-limited cumulative summation to this cumulative causal input and maps it to the [0,1] interval via a nonlinear compression function, which serves as the updated activation membership degree of that node in this iteration. Taking the hyperbolic tangent function tanh(·) as the compression function as an example, its calculation method can be expressed as: Let x be the cumulative causal input received by node j after the previous iteration. j Then, the membership degree of this node in this round of update activation is taken as [tanh(β·x)]. j )+1] / 2, where β is the steepness control parameter; an upper limit threshold is also set, and the calculated value is truncated to the threshold when it exceeds the threshold. The essence of this design is to superimpose the combined influence of multiple upstream nodes according to causal logic, and simulate the psychological law of diminishing marginal effects in reality through nonlinear transformation, that is, when a certain risk factor is already at a high level, the marginal deterioration effect will gradually weaken when other risk factors continue to exert the same degree of influence. After each iteration, the system calculates the maximum absolute change in the activation membership degree of each node between the two rounds. If this value is lower than the preset convergence tolerance (e.g., 0.001), the inference is determined to have reached a steady state, and the final activation membership degree of each concept node is output.

[0057] After convergence, the system marks concept nodes whose activation membership exceeds a preset vulnerability threshold (e.g., 0.7) as potential vulnerable nodes. For each potential vulnerable node, the system uses that node as the starting point and performs a depth-first path search along all reachable causal paths originating from that node, traversing all paths that start from that node. For each complete path, the system calculates the product of the activation membership increment between adjacent nodes on the path and the edge weight, and multiplies the products of all edges on the path to obtain the chain reaction strength of that path. The physical meaning of this value is the degree to which the current potential vulnerable node, if further activated, will ultimately affect the state of the terminal node through this causal path. The system then outputs multiple chain risk paths triggered by each potential vulnerable node, sorted from highest to lowest chain reaction strength.

[0058] Based on the ranking results, the system reverse-engineers concept nodes located in the middle of the chain of risk paths with the highest chain influence intensity and whose incoming edge weights in the causal weighting matrix are adjustable from these paths, marking them as intermediate controllable nodes. "Adjustable" means that the activation state of such nodes can be artificially intervened and adjusted through dialogue guidance and cognitive restructuring, such as concept nodes corresponding to the use of emotion regulation skills and positive social behaviors. For each intermediate controllable node, the system automatically generates a preventative dialogue node. This node contains intervention topic text aimed at reducing the activation membership of the intermediate controllable node, and a directional marker indicating the generation goal (e.g., the direction of activation value reduction). The intervention topic text is automatically generated by the large language model based on pre-set cognitive behavioral strategy templates and node semantic labels. Examples include "guiding individuals to recall experiences of successfully handling similar emotions" and "inviting individuals to discuss available social support." Finally, the system injects the generated series of preventative dialogue nodes, in order of path risk priority, into the context of the responses to be generated in low-risk retrieval-based multi-turn dialogue units in the form of flexible guiding statements. The injection method involves concatenating the intervention topic text of the current preventative dialogue node at the head of the queue into the implicit guidance instruction area of ​​the role prompts after obtaining candidate reference dialogue fragments in the retrieval enhancement generation process and before calling the large language model to generate the final response. In this way, when the large language model generates a personalized response, it will naturally integrate the intervention topic into the dialogue, buffering and guiding potentially vulnerable nodes in advance without disrupting the fluency of the dialogue, thus reducing the possibility of cascading risks. Through this preferred embodiment, the system upgrades the monitoring of discrete risk indicators to continuous, deductive causal network dynamic analysis, achieving a leap from "post-event risk discovery" to "pre-event prediction and proactive prevention of cascading risks," further enhancing the proactive protection capabilities of the psychological support agent in long-term companionship interventions.

[0059] In summary, this invention achieves accurate identification, tiered intervention, and safe closed-loop management of adolescents with psychological distress through a two-level hybrid architecture consisting of automated clinical scale screening and machine learning prediction models. This is combined with structured high-risk assessment based on rules and large language models, retrieval-enhanced generative low-risk dialogue, full-time safety monitoring, and constrained referral recommendations. This effectively improves the safety, professionalism, and scalability of large language models in mental health application scenarios.

[0060] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made using the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-level risk-stratified intelligent agent system for adolescent psychological support, characterized in that, include: The system includes a risk stratification module, a stratified intervention module, a safety and dynamic monitoring module, and a professional referral module. The risk stratification module is used to perform hybrid architecture stratification, including: An automated clinical scale screening unit is used to receive the BSI-53 scale completed by an individual and calculate the total score. When the total score is greater than or equal to a preset clinical threshold, the professional referral module is triggered. When the total score is less than the preset clinical threshold, the machine learning-based risk prediction unit is triggered. The risk prediction unit is used to extract six feature variables from a preset scale filled in by an individual, input them into a pre-trained BPNN model after standardization, output risk probability values ​​and compare them with a preset classification threshold to classify individuals as high-risk or low-risk. The hierarchical intervention module includes a high-risk embedded question-and-answer unit and a low-risk retrieval-based multi-turn dialogue unit. The high-risk embedded question-answering unit is used to conduct natural language assessments on high-risk individuals using a rule-based predefined dialogue framework, focusing on suicide risk, self-harm behavior, and somatization symptoms. It then calls a large language model to generate supportive responses and escalation judgments, and triggers manual review or redirection to the low-risk retrieval-based multi-turn dialogue unit based on the judgment results. The low-risk retrieval-based multi-turn dialogue unit is used to conduct multi-turn dialogues with low-risk individuals using a retrieval-enhanced generation method. Based on the individual's current intervention stage and input content, it retrieves similar historical dialogue records from a knowledge base that includes peer mentor dialogues and public psychological counseling dialogues, and generates personalized responses. The security and dynamic monitoring module is used to acquire multi-turn dialogue context in real time, call the large language model to perform risk classification through preset risk judgment prompt words, and trigger manual review when any risk level is identified. The professional referral module is used to collect individual preferences through multi-round dialogue when a referral trigger signal is received, and to perform hierarchical recall and constraint sorting in the referral resource database to output recommended professional psychological assistance resources.

2. The system according to claim 1, characterized in that, The preset clinical threshold is 63 points.

3. The system according to claim 1, characterized in that, In the risk prediction unit, the six feature variables are derived from subdomain scores of the Life Events Self-Rating Scale, the Perceived Social Support Scale, and the Emotional Intelligence Scale; the BPNN model is a single-hidden-layer feedforward neural network with six features in the input layer, three neurons in the hidden layer using the ReLU activation function, and the output layer using the Sigmoid function to output the risk probability. The standardization is based on the parameters of the training dataset; the preset classification threshold is 0.

5.

4. The system according to claim 1, characterized in that, The specific rules for evaluation in the high-risk embedded question-answering unit are as follows: First, a suicide risk screening is conducted. If an individual expresses suicidal ideation or behavior, a manual review is triggered directly. Otherwise, a self-harm behavior screening will be conducted, and if self-harm behavior is found, a manual review will be triggered. Otherwise, a somatization assessment is performed, which includes four progressive questions Q1, Q2, Q3, and Q4, where Q1 addresses somatization symptoms, and Q2, Q3, and Q4 address anxiety and depression symptoms, depression symptoms, and anxiety symptoms, respectively. When Q1 is positive and any one of Q2, Q3 or Q4 is positive, manual review is triggered; Manual review is triggered when Q1 is negative and Q2, Q3 and Q4 are all positive. When all four rounds of Q1 are negative, the individual is returned to the low-risk retrieval multi-turn dialogue unit.

5. The system according to claim 1, characterized in that, In the low-risk retrieval-based multi-turn dialogue unit, the knowledge base construction and retrieval process includes: Clean up counseling dialogue records and make public the adolescent psychological counseling dialogue dataset, remove cases containing clear mental illness diagnoses and self-harm and suicide content, and categorize them by intervention stage and theme; The dialogue text is vectorized and stored in a vector database, with individual statements as the core and adjacent sentences combined to form retrieval units. During the search, the peer mentor dialogue database is searched according to the individual's current intervention stage. If the results are insufficient, the individual's input is subject to topic identification and supplemented by searching the public dataset of the corresponding topic. The current individual input and candidate dialogue fragments are vector-encoded, and the cosine similarity score is calculated. After sorting in descending order and removing duplicates, the top 3 are selected as the reference for generating the response.

6. The system according to claim 5, characterized in that, The low-risk retrieval-based multi-turn dialogue unit also includes a behavioral task generation constraint mechanism: During the first consultation each week, at least one behavioral task is generated based on the comprehensive analysis of multiple rounds of dialogue, and the task content is written into the individual's long-term memory module. Within the same intervention period, new behavioral tasks should not be generated repeatedly. Instead, historical task information should be used for tracking, feedback, and encouragement. When the task is detected to be completed or the next intervention cycle is entered, the task generation permission is re-enabled.

7. The system according to claim 1, characterized in that, The security and dynamic monitoring module is specifically used for: Extract the most recent k=6 rounds of interaction between the individual and the agent as a context window. Input the context and predefined risk classification criteria into the large language model for structured reasoning and output the risk level L1, L2 or L3. L1 indicates extremely negative emotions but no clear intention to self-harm or commit suicide. L2 indicates the existence of suicide or self-harm risk but has not entered the execution stage. L3 indicates the existence of immediate life danger or the implementation of self-harm or suicide behavior. When any risk level is identified, the agent's response is paused and a manual review is triggered. An expert will assess the situation based on the complete dialogue context. If the risk is manageable, the interaction will resume; otherwise, the professional referral module or emergency intervention will be triggered.

8. The system according to claim 1, characterized in that, The professional referral module includes a referral resource database, which stores manually verified information on psychological service hotlines and institutions; the referral trigger signals include: the BSI-53 total score is greater than or equal to the preset clinical threshold, or the individual is determined by manual review in the safety and dynamic monitoring module to be unsuitable for further intervention; The multi-turn dialogue collects individual preferences including population type, regional information, and cost preferences; The hierarchical recall is carried out in the following order of priority: joint provincial and municipal matching, city matching, province matching, provincial record matching, and national resource matching. When the matching fails, the city field is subjected to suffix removal and normalization as a fallback match. After recalling the candidate list, the candidate list is reordered with the large language model constrained by individual preference inputs, allowing selection only from the candidate list, and outputting structured information of up to 5 matching institutions.

9. A multi-level risk stratification method for adolescent psychological support, applied to the system described in any one of claims 1 to 8, characterized in that, include: Receive the completed BSI-53 scale from the individual, calculate the total score, and refer the individual to professional services if the total score is greater than or equal to the preset clinical threshold; otherwise, proceed to the next step. Six feature variables are extracted from the pre-defined scale filled in by the individual, standardized and input into the pre-trained BPNN model to obtain the risk probability value and compare it with the threshold. If it is lower than the threshold, the low-risk retrieval multi-turn dialogue unit is used for intervention; otherwise, the high-risk embedded question-answering unit is used for intervention. During the intervention process using the aforementioned low-risk retrieval-based multi-turn dialogue unit, personalized responses are generated by combining the knowledge base and historical dialogues, and dialogue risks are monitored in real time. During the intervention using the high-risk embedded question-answering unit, suicide risk, self-harm behavior and somatization symptoms are assessed sequentially through natural language dialogue. Based on the large language model, it is determined whether to escalate and trigger manual review. Throughout the intervention process, the context of multi-turn dialogues is classified in real time for risk assessment. When a risk is identified, manual review is triggered, and referral or resumption of interaction is performed based on the review results. When the referral criteria are met, individual preferences are collected and hierarchical recall and constrained ranking are performed in the referral resource database to output recommended professional institutions.

10. The method according to claim 9, characterized in that, During the intervention process of the high-risk embedded question-and-answer unit, the somatization symptom assessment uses four progressive questions. If the first question is positive, any subsequent question being positive will trigger manual review. If the first question is negative, all subsequent questions must be positive to trigger manual review. If all questions are negative, the process returns to the low-risk retrieval-based multi-turn dialogue unit.

Images