Mental health analysis and intervention method based on multi-Agent collaborative evaluation

Abstract

The invention provides a psychological health analysis and intervention method based on multi-Agent collaborative assessment, which comprises the following steps of: shunting information and dialogue contents input by a user through an initial shunting Agent module, and then performing parallel and independent multi-dimensional psychological assessment by using a plurality of different assessment Agent modules. Meanwhile, each assessment Agent module corresponds to a psychological dimension and is provided with a special vertical domain knowledge base, then, structured output results of the assessment Agent modules are received and integrated through the central coordination and treatment agent, a collaborative and unified psychological health assessment report is generated, and finally, based on the assessment report, a psychological health assessment report is established. The treatment intervention Agent module is used for generating a personalized mental health intervention scheme matched with the user, and a precise intervention scheme is provided for the user.

Description

Technical Field

[0001] This invention relates to the field of mental health management technology, specifically to a mental health analysis and intervention method based on multi-agent collaborative assessment. Background Technology

[0002] The core content of the field of mental health includes mental health information collection, mental health risk assessment, and mental health intervention. In recent years, with the continuous development of society, more and more attention has been paid to the preventive effect of mental health education for college students. Mental health is the first line of defense against mental illnesses such as depression and anxiety. Early intervention can reduce the incidence of crisis events, enable college students to become qualified members of society with empathy and cooperation, and thus promote college students' self-awareness and emotional management, and cultivate sound personality and social responsibility.

[0003] For example, Chinese patent application No. 202510800276.6, published on September 16, 2025, discloses a college student mental health service system and method based on AI Agent. The system includes a mental health measurement module, a social personality profiling module, and a mental health intervention module. The measurement agent uses a standardized mental health scale to measure the current mental health status of college students from a horizontal perspective; the profiling agent uses social media data of college student users obtained through web crawling to analyze their historical mental health status and MBTI personality profiles from a vertical perspective; based on the mental health scale data and social media data, the system adaptively recommends counselor agents to college students. The counselor agents are externally connected to psychological counseling expert guides and real counseling case data to provide college students with professional and effective mental health intervention services, applicable to scenarios such as emotional support for college students and university decision-making assistance.

[0004] The aforementioned literature utilizes a psychological health measurement module, a social personality profiling module, and a psychological health intervention module, combined with psychological health scales, social media data, and expert guidelines for psychological counseling, to generate multidimensional psychological health scores, descriptive statistical reports, sentiment analysis reports, and adaptive psychological health intervention services. It primarily combines psychological health data and social media data to recommend counselors. However, it does not rigorously analyze the user's current behavior, emotions, and cognition, especially their cognitive state. Furthermore, after assessment, it cannot recommend different intervention plans based on the user's current state. For example, for those with relatively high cognitive levels, initial treatment through cognitive intervention is needed to restore the user's cognition to a stable state before timely intervention can be effectively implemented. Therefore, because it does not conduct separate vertical assessments of user data through different assessment modules and does not coordinate and unify the results of these vertical assessments through a central coordination module, it cannot accurately recommend precise intervention plans for users. Summary of the Invention

[0005] The purpose of this invention is to provide a mental health analysis and intervention method based on multi-agent collaborative assessment. By adopting a collaborative architecture of "multi-agent vertical assessment + central coordination intervention", it provides users with precise intervention plans.

[0006] To achieve the above objectives, this invention provides a method for mental health analysis and intervention based on multi-agent collaborative assessment, implemented through a mental health analysis and intervention system; including the following steps: S1 acquires the user identity information and corresponding input information of the user entering the mental health analysis and intervention system, and performs preprocessing through the initial triage Agent module; S2 evaluates user questions based on behavior, cognition, and emotion after initial triage, obtaining corresponding evaluation values. It then performs in-depth analysis on the vertical domain-specific RAG knowledge base that matches the behavior, cognition, and emotion, and sums the results with the corresponding evaluation values ​​according to the preset weights of behavior, cognition, and emotion to obtain a comprehensive structured evaluation result. In step S2, the S3 central coordination and treatment agent receives the integrated structured assessment results and the assessments of behavior, cognition, and emotion to obtain the corresponding assessment values. Based on the assessment results and the assessments of behavior, cognition, and emotion, the corresponding assessment values ​​are used to recommend intervention priorities in the multi-dimensional assessment report corresponding to the generated user identity information. S4 uses the treatment intervention agent module to retrieve data from the intervention measures RAG knowledge base and form a matching intervention strategy, and then forms a corresponding intervention plan; S5 dynamically adjusts the weights corresponding to behavior, cognition, and emotion based on the post-intervention situation, and performs adaptive intervention for users in short-term and long-term dialogue interactions.

[0007] The above method employs a multi-dimensional intelligent agent vertical domain architecture, equipping each different evaluation agent module with a dedicated RAG knowledge base. It assesses the user's current situation from three aspects: emotion, cognition, and behavior, obtaining evaluation values. These three evaluation values ​​are then weighted and summed with preset weights to obtain a comprehensive structured result. An intervention plan is selected based on this comprehensive structured result and the three evaluation values. This approach combines the comprehensive structured structure with intuitive judgment based on the three evaluation values. When a problem arises with a crucial emotion evaluation value, an emotion-based adjustment intervention plan is directly initiated. Similarly, when cognitive and behavioral evaluation values ​​are high, cognitive and behavioral interventions are also considered. After implementing an intervention plan, the emotion, cognition, and behavioral evaluation values ​​are re-evaluated, and adjustments to the short-term and long-term dialogue intervention plans are determined based on these values. This ensures adaptive adjustment to a normal state after adjustments, preventing over-intervention that could lead to inaccurate interventions.

[0008] Furthermore, in step S1, the user connects to an external identity authentication system through an identity authentication interface to achieve secure and independent login to the mental health analysis and intervention system; In step S1, after the user logs into the mental health analysis and intervention system, and with the user's authorization and in compliance with the security and privacy protection module, the mental health analysis and intervention system exchanges non-anonymized information with the mental health counseling center management system that matches the external identity authentication system. In step S1, after preprocessing by the initial triage Agent module, the process also includes: performing initial triage for intent identification; if the user is assessed as being at high risk, the crisis protocol of the mental health analysis and intervention system is triggered, and the mental health analysis and intervention system sends encrypted alert information to the mental health counseling center management system or the user's emergency contact via API call, SMS, or email; otherwise, proceed to step S2.

[0009] The above settings can assess whether the risk is severe based on the information data entered by the user, and then identify and notify the external systems or contacts associated with the user in advance, so as to take timely human intervention for the user.

[0010] Furthermore, step S1 also includes anonymizing the user input data through a security and privacy protection module: encrypting data transmission and static data, and after the user completes the dialogue, only the desensitized structured assessment data is retained in the mental health analysis and intervention system for long-term trend analysis, while the non-desensitized structured assessment data is automatically deleted.

[0011] The above settings ensure the confidentiality of data generated by users in the mental health analysis and intervention system, and prevent data leakage.

[0012] Furthermore, step S2 also includes: steps S2.1 to S2.3. When a user is assessed as not at high risk, S2.1 engages in dialogue with the user through a Large Language Model (LLM) and generates an initial response. Then, based on a vertical domain-specific RAG knowledge base matched with different assessment agent modules, it independently retrieves the data information from the user's response and generates structured assessment results for the corresponding behavioral, cognitive, and emotional dimensions. S2.2 Based on the evaluation values ​​from the structured evaluation results of three different dimensions, three different corresponding weights W_b, W_c, and W_e are matched, and the sum of the three different weights is 1. Then, the comprehensive evaluation value S_base is calculated using formula (1). S_base = (Score_e * W_e) + (Score_c * W_c) + (Score_b * W_b) (1), Score_e represents the evaluation value of the structured assessment result in the emotion dimension, and W_e represents the weight corresponding to the emotion dimension, with a value range of 0.3 - 0.4; Score_c represents the evaluation value of the structured assessment result in the cognition dimension, and W_c represents the weight corresponding to the cognition dimension, with a value range of 0.3 - 0.4; Score_b represents the evaluation value of the structured assessment result in the behavior dimension, and W_b represents the weight corresponding to the behavior dimension, with a value range of 0.3 - 0.4. S2.3 By using a Large Language Model (LLM) to integrate structured assessment results, personalized responses that match the user are generated, and corresponding personalized intervention plans are provided to the user.

[0013] The above settings, through the combination of a large language model (LLM) and a vertical domain-specific RAG knowledge base that matches different assessment agent modules, can accurately respond to users' dialogue data and provide intervention paths, thereby achieving precise guidance for users' mental health.

[0014] Furthermore, step S2.2 also includes: steps S2.2.1 to S2.2.3. S2.2.1 If the evaluation value of any dimension in the three different dimensions of the structured evaluation results is >9, the comprehensive evaluation value S_base will become invalid and the crisis protocol will be triggered. S2.2.2 If the assessment value of the structured assessment result in the emotion dimension is >7 and the assessment value of the structured assessment result in the cognition dimension is >7, the resulting comprehensive evaluation result is a high-risk level, and a cognitive behavioral therapy (CBT) intervention plan is provided to the user in the resulting assessment report. S2.2.3 If the score of the structured assessment result in the emotion dimension is >7 and the score of the structured assessment result in the behavior dimension is <4, then the user will be provided with an intervention path for behavior activation in the resulting assessment report.

[0015] If any of the above settings are assessed at a value greater than 9, they can be adjusted in a timely manner through a crisis protocol. When emotions and cognition are excessively high, cognitive behavioral therapy can be used to intervene in the behavior and determine whether emotional and cognitive adjustments can be achieved.

[0016] Furthermore, step S3 also includes: The psychological profile is dynamically updated in real time based on the dynamic data during the initial triage process of the Agent module in step S1, the process of generating personalized responses for users by integrating the large language model (LLM) in step S2, and the process of users selecting the appropriate personalized intervention path.

[0017] The above settings allow for dynamic updates to the psychological profile based on data from different stages, making the psychological profile more closely match the user's actual situation and enabling the user to better understand their own psychological state.

[0018] Furthermore, the multi-dimensional evaluation report in step S3 includes emotional state, cognitive pattern, and behavioral performance, as well as a comprehensive evaluation value S_base calculated from the evaluation scores corresponding to emotional state, cognitive pattern, and behavioral performance.

[0019] The above settings enable users to take corresponding intervention measures through multi-dimensional assessment reports, and to intervene in users' psychological problems in a timely manner.

[0020] Furthermore, step S4 includes S4.1 to S4.3. The S4.1 Treatment Intervention Agent module retrieves corresponding emotional states, cognitive patterns, and behavioral performances from the intervention measures RAG knowledge base based on the multi-dimensional assessment report, and recommends corresponding intervention strategies through immediate intervention goals corresponding to emotional states, core intervention goals corresponding to cognitive patterns, and supportive intervention goals corresponding to behavioral performances. S4.2 In the cognitive pattern retrieval process, the user's verbal responses are mapped to standardized cognitive distortion types in cognitive behavioral therapy (CBT) to obtain different degrees of cognitive bias. S4.3 Based on different degrees of cognitive bias, corresponding intervention strategies are formed, and personalized intervention plans are output to users by combining the intervention strategies corresponding to the immediate intervention goals and the intervention strategies corresponding to the supporting intervention goals.

[0021] The above settings enable the system to dynamically combine data retrieval from the knowledge base storage module with psychological profiling to accurately provide intervention strategies for different levels of risk based on the actual psychological state of different users, thereby providing effective solutions and interventions for the psychological problems that users have.

[0022] Furthermore, step S4.2 includes steps S4.2.1 to S4.2.3. S 4.2.1 The cognitive assessment agent module performs word segmentation and syntactic analysis on the user's input language response to extract the core viewpoints, beliefs or predictions expressed by the user; S 4.2.2 will extract the core viewpoints, beliefs or predictions expressed by users and search and match them with the RAG knowledge base of specific intervention measures; S 4.2.3 Compare the semantic vectors of the user’s core viewpoints, beliefs or predictions with the semantic vectors of example patterns in the RAG knowledge base of the specific intervention measures, and then use the numerical value of the similarity to judge the different degrees of cognitive bias.

[0023] The above settings, achieved through word segmentation and extraction of user-input language during the cognitive retrieval process, ensure the reliability of cognitive assessment.

[0024] Furthermore, step S5 also includes steps S5.1 to S5.3. During short-term dialogue interactions, S5.1 users are evaluated through the Emotion Assessment Agent, Cognitive Assessment Agent, and Behavioral Assessment Agent modules, and corresponding weights are assigned. After the crisis intervention mode is triggered, the corresponding weight ratio is reduced so that the weight ratio corresponding to the emotion assessment value is equal to the sum of the weight ratio corresponding to the cognitive assessment value and the weight ratio corresponding to the behavioral assessment value, and the user is guided to implement the corresponding intervention plan. After intervention, a reassessment is conducted until the emotional crisis is resolved. The weighting ratio is then readjusted, with cognition-behavior weighting taking precedence, to guide users in cognitive restructuring and behavioral planning. S5.2 For users with long-term dialogue interactions, the weights corresponding to the behaviors, cognitions, and emotions that have the best effects in historical data are adjusted. If it is detected that the evaluation value of different adaptive strategies decreases in subsequent dialogue interactions, the weights corresponding to the original behaviors, cognitions, and emotions are adjusted back.

[0025] The above settings can evaluate users in different conversations, and then adjust the weight ratio of the corresponding adaptive strategy after evaluation. This allows for adjustments to the user's focus from different perspectives to adapt to different user states, thereby improving the psychological guidance effect. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall architecture of the present invention.

[0027] Figure 2 This is a flowchart of the psychological dialogue based on LLM+RAG in this invention.

[0028] Figure 3 This is a logic diagram for recommending intervention paths in this invention.

[0029] Figure 4 This is a schematic diagram illustrating the long-term interactive adaptive strategy adjustment for user groups in this invention.

[0030] Figure 5 This is a schematic diagram of real-time strategy adjustment in this invention. Detailed Implementation

[0031] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0032] like Figure 1-5 As shown, a mental health analysis and intervention method based on multi-agent collaborative assessment is implemented through a mental health analysis and intervention system. The mental health analysis and intervention system includes a user interface module, a security and privacy protection module, a multi-dimensional assessment intelligent agent group, a central coordination and treatment intelligent agent, a knowledge base storage module, and an external system interface module. The user interface module includes an identity authentication interface; the external system interface module includes a data exchange interface and an emergency referral interface; the knowledge base storage module has a data storage sub-module to ensure the security of stored data.

[0033] In this embodiment, the multi-dimensional assessment intelligent agent group includes an emotion assessment agent module, a cognitive assessment agent module, and a behavior assessment agent module. The central coordination and treatment intelligent agent includes a central coordination agent module. The different assessment agent modules in the multi-dimensional assessment intelligent agent group communicate with the initial distribution agent module and the central coordination agent module through internal interfaces, respectively. At the same time, the initial distribution agent module is connected to the external system interface module, and the central coordination agent module is connected to the knowledge base storage module. This facilitates efficient communication between different modules and integration with the external ecosystem through the external system interface module.

[0034] In this embodiment, taking university students as the user and the university's unified identity authentication system as the external identity authentication system as the example, the following specific steps are also included: S1 users connect to the university's unified identity authentication system through the identity authentication interface, avoiding the duplication of user system construction and enabling secure, individual login to the mental health analysis and intervention system. Then, under the premise of user authorization and compliance with the security and privacy protection module conditions, the mental health analysis and intervention system exchanges non-anonymized information with the university's psychological counseling center management system, which is matched with the university's unified identity authentication system. If the mental health analysis and intervention system identifies a user as high-risk, it automatically generates an encrypted "to be monitored" work order and pushes it to the counselor's workbench of the university's psychological counseling center management system. Next, the user input data is anonymized through the security and privacy protection module. In this embodiment, this includes encrypting data transmission (TLS 1.3+) and static data (AES-256 encryption). After the user completes the dialogue, only the de-anonymized structured assessment data is retained in the mental health analysis and intervention system for long-term trend analysis, while the non-de-anonymized structured assessment data is automatically deleted.

[0035] After anonymizing the data, the user's dialogue content is preprocessed and intent is identified through the initial triage Agent module. If the user is assessed as being at high risk, such as when the emotion recognition module of the mental health analysis and intervention system identifies keywords such as "tired," "leaving," and "relaxed" in the user's questions and answers, it is judged as being in a state of "severe anxiety." The risk assessment module of the mental health analysis and intervention system, based on recent usage records and historical emotion fluctuation curves, automatically marks the college student's status as "severe risk" and triggers the crisis protocol of the mental health analysis and intervention system. The mental health analysis and intervention system sends encrypted alert information to the university's psychological counseling center management system or the user's emergency contact by calling the API gateway, SMS, or email gateway. In the subsequent tracking process, the mental health analysis and intervention system continuously pushes three warm reminders and safety care greetings to the college student to ensure the continuity of psychological support; otherwise, proceed to step S2.

[0036] S2 as Figure 1 and 2 As shown, different evaluation agent modules in the multi-dimensional evaluation intelligent agent group conduct different evaluations on user problems after being diverted by the initial diversion agent module. In this embodiment, when a user is evaluated as not severely at risk, if the emotion recognition module of the mental health analysis and intervention system identifies keywords such as "anxiety" and "insomnia" in the user's problem and answer, it is judged as a "moderate anxiety" state, and the risk assessment module of the mental health analysis and intervention system judges it as moderate risk. Then, the intelligent dialogue module of the Large Language Model (LLM) engages in dialogue with the user based on humanistic guidance, providing empathetic comfort and generating initial responses, such as, "It's already a very important step that you can proactively express your feelings. We're here to help you further explore these feelings," while guiding students to be aware of and express their emotions; proceeding to step S2.1, S2.1 independently retrieves user-response data from vertical domain-specific RAG knowledge bases matched with the emotion assessment agent module, cognitive assessment agent module, and behavioral assessment agent module, and generates structured assessment results for the corresponding behavioral, cognitive, and emotional dimensions. Specifically, upon receiving a request, each assessment agent module executes its own "retrieval-enhancement-generation" process and returns the structured assessment results in "JSON" format to the central coordination agent module. The agent module, in artificial intelligence, is used to identify, analyze, and respond to user emotional states. Its implementation can be achieved through dictionary-based methods, machine learning methods, and deep learning methods. Dictionary-based methods calculate text sentiment scores using a predefined sentiment dictionary (e.g., negative / positive words). If words detected through word segmentation are compared with words in the predefined sentiment dictionary to determine whether they are negative or positive, the current emotional state is determined. Machine learning methods utilize algorithms such as Naive Bayes and Support Vector Machines to learn sentiment features from labeled data. Deep learning methods employ models such as CNN and LSTM to automatically extract deep text features and are adept at capturing contextual dependencies. The implementation method of the emotion assessment agent module is existing technology and will not be elaborated upon here.

[0037] The cognitive assessment agent module and the emotion assessment agent module are implemented in the same way, only the cognitive keywords used are different.

[0038] The Behavior Assessment Agent module and the Emotion Assessment Agent module are implemented in the same way, only the cognitive keywords used are different.

[0039] S2.2 Based on the evaluation values ​​from the structured evaluation results of three different dimensions, three different corresponding weights W_b, W_c, and W_e are matched, and the sum of the three different weights is 1. Then, the comprehensive score S_base is calculated using formula (1). S_base = (Score_e * W_e) + (Score_c * W_c) + (Score_b * W_b) (1), Score_e represents the evaluation value of the structured assessment result in the emotion dimension, and W_e represents the weight corresponding to the emotion dimension, with a value range of 0.3 - 0.4; Score_c represents the evaluation value of the structured assessment result in the cognition dimension, and W_c represents the weight corresponding to the cognition dimension, with a value range of 0.3 - 0.4; Score_b represents the evaluation value of the structured assessment result in the behavior dimension, and W_b represents the weight corresponding to the behavior dimension, with a value range of 0.3 - 0.4; as shown in Table 1 below. Table 1

[0040] Then proceed to step S2.2.1. S2.2.1 If the evaluation value of any dimension in the three different dimensions of the structured evaluation results is >9, the comprehensive score S_base will be invalidated and the crisis protocol will be triggered; S2.2.2 If the assessment value of the structured assessment result in the emotion dimension is >7 and the assessment value of the structured assessment result in the cognition dimension is >7, the resulting comprehensive evaluation result is a high-risk level, and the user is provided with an intervention path of cognitive behavioral therapy (CBT) in the resulting assessment report. S2.2.3 If the evaluation value of the structured assessment result in the emotion dimension is >7 and the evaluation value of the structured assessment result in the behavior dimension is <4, then the user will be provided with an intervention path for behavioral activation in the generated assessment report. Next, the Large Language Model (LLM) is used to generate personalized responses that match the user based on the structured assessment results. At the same time, it provides the user with corresponding personalized intervention paths, such as "recommending a 3-minute breathing training audio and guiding the college student to make an appointment for offline human consultation to obtain more in-depth professional help." In this way, by combining the Large Language Model (LLM) with a vertical domain-specific RAG knowledge base that matches different assessment agent modules, it is possible to accurately respond to the user's dialogue data and provide intervention paths, thereby achieving precise guidance for the user's mental health.

[0041] S3 as Figure 2 As shown, during the initial triage process by the Agent module in step S1, the process of generating personalized responses for users through Large Language Model (LLM) fusion in step S2, and the process of users selecting the corresponding personalized intervention path, the dynamic data is updated in real time to dynamically update the user's initial preset psychological profile. Then, the central coordination and treatment agent receives the evaluation results from step S2 and generates a multi-dimensional evaluation report according to the preset integration rules. In this embodiment, the multi-dimensional evaluation report includes emotional state, cognitive pattern, and behavioral performance, as well as a comprehensive evaluation value S_base calculated from the evaluation scores corresponding to emotional state, cognitive pattern, and behavioral performance, and the corresponding scale chart.

[0042] S4, through the treatment intervention agent module, retrieves data from the intervention measures RAG knowledge base stored in the knowledge base storage module to form a matching intervention strategy, and then outputs a personalized intervention plan to the user. Specifically, this includes steps S4.1 to S4.3. The S4.1 Treatment Intervention Agent module retrieves corresponding emotional states, cognitive patterns, and behavioral performances from the intervention measures RAG knowledge base based on the multi-dimensional assessment report, and recommends corresponding intervention strategies through immediate intervention goals corresponding to emotional states, core intervention goals corresponding to cognitive patterns, and supportive intervention goals corresponding to behavioral performances. S4.2 In the cognitive pattern retrieval process, the user's verbal responses are mapped to standardized cognitive distortion types in cognitive behavioral therapy (CBT) to obtain different degrees of cognitive bias, including steps S4.2.1 to S4.2.3. S 4.2.1 uses the cognitive assessment agent module to perform word segmentation and syntactic analysis on the user's input language responses, extracting the core viewpoints, beliefs, or predictions expressed by the user. For example, if the user inputs: "I have my defense tomorrow, and I'm not prepared at all. I think I'm definitely going to mess it up and won't be able to graduate," the key expressions extracted are: Statement 1: "I was not prepared at all." Statement 2: "I think I'm definitely going to mess this up." Statement 3: "Then I couldn't graduate." S 4.2.2 will extract the core viewpoints, beliefs, or predictions expressed by the user and search and match them with the RAG knowledge base for specific intervention measures. In this embodiment, based on the extracted statements 1, 2, and 3, The first type of distortion is catastrophizing, which is defined as ignoring the more likely outcomes and instead imagining and the worst possible scenario; example patterns are: "If...then we're doomed", "It's certain that...", "This will be a disaster"; Distortion type two is all-or-nothing thinking, defined as viewing things in black-and-white binary categories, with no middle ground; example patterns: "not at all...", "completely...", "either success or failure"; S 4.2.3 Compare the similarity between the semantic vectors of the user's core viewpoints, beliefs, or predictions and the semantic vectors of the "example patterns" in the RAG knowledge base for specific interventions. In the extracted statement 1, "I am not prepared at all" is highly similar to the example pattern of "all or nothing", with a similarity score of 0.92; In the extracted statement 2, "I think I'm definitely going to mess up" is highly similar to the example patterns of "catastrophizing" and "negative prediction", with a similarity score of 0.88; In the extracted statement 3, "then unable to graduate" is highly similar to the definition of "catastrophizing"; Then, the degree of cognitive bias is judged by the numerical similarity score; S4.3 Based on different degrees of cognitive bias, corresponding intervention strategies are formed, and personalized intervention plans are output to users by combining the intervention strategies corresponding to the immediate intervention goals and the intervention strategies corresponding to the supporting intervention goals. In this embodiment, if the emotion assessment score is >7 and the cognitive bias is of the "catastrophizing" type, and the behavior assessment score is... 5. This triggers the "cognitive restructuring priority" intervention strategy. If the emotional assessment score is >7, the cognitive bias is of the "catastrophizing" type, and the behavioral assessment score is >6, then the "behavioral activation priority" intervention strategy is triggered. This allows for the creation of personalized intervention plans based on different assessment states, which are then delivered to the user.

[0043] like Figure 5 As shown, adaptive strategies are dynamically adjusted for users in short-term and long-term dialogue interactions by adjusting the weight ratios corresponding to behavioral assessment values, cognitive assessment values, and emotional assessment values, respectively; specifically, steps S5.1 to S5.3 are included. During short-term dialogue interactions, S5.1 users are evaluated through the Emotion Assessment Agent module, Cognitive Assessment Agent module, and Behavioral Assessment Agent module, and corresponding weights are assigned. After the crisis intervention mode is triggered, the corresponding weight ratio is reduced so that the weight ratio corresponding to the emotion assessment value is equal to the sum of the weight ratio corresponding to the cognitive assessment value and the weight ratio corresponding to the behavioral assessment value, and the user is guided to take corresponding intervention measures. After intervention, a reassessment is conducted until the emotional crisis is resolved. The weighting ratio is then readjusted, with cognition-behavior weighting taking precedence, to guide users in cognitive restructuring and behavioral planning. S5.2 For users with long-term dialogue interactions, the weights corresponding to the behaviors, cognitions, and emotions that have the best effects in historical data are adjusted. If it is detected that the evaluation value of different adaptive strategies decreases in subsequent dialogue interactions, the weights corresponding to the original behaviors, cognitions, and emotions are adjusted back.

[0044] In one embodiment, for a user's initial state during a short conversation: Emotion = 9, Cognition = 8, Behavior = 7 → "Crisis Intervention Mode" is triggered. At this time, the fusion rules prioritize "Safety and Stability," and the weights are temporarily adjusted to: Emotion (0.5), Cognition (0.3), Behavior (0.2). The goal is to quickly de-escalate the emotional storm.

[0045] After intervention: The system guided the user through deep breathing and grounding exercises, and the user reported feeling "a little better." The system then conducted a quick assessment, which yielded the following results: Emotion = 6, Cognition = 7, Behavior = 7.

[0046] Integration Strategy Adjustment: The emotional crisis is temporarily resolved, and the core issues (high cognitive bias and avoidance behavior) emerge. The system automatically switches the integration weights to a "cognitive-behavior-dominated mode": emotion (0.3), cognition (0.4), behavior (0.3).

[0047] For users who engage in long-term conversations, those rated as "high anxiety, high cognitive bias, and low motivation" during a single conversation are found by the system through historical data to have the best long-term efficacy when using a fusion strategy of "cognition (0.45) - behavior (0.35) - emotion (0.20)" (i.e., first addressing erroneous thinking, then developing small-step action plans, which naturally alleviates emotions). Therefore, the system will personalize this fusion weight for such users and intervene accordingly.

[0048] In this embodiment, step 2 of the RAG knowledge base also includes the following processing: (a) Data collection and cleaning: collect authoritative textbooks, treatment guidelines, academic papers and structured scales (such as SCL-90, PHQ-9) in the field of mental health to form the original corpus.

[0049] (b) Text preprocessing and chunking: The corpus is cleaned and deduplicated. A recursive character text segmenter is used to divide long texts into appropriately sized chunks based on semantic relevance and to set overlapping areas to ensure contextual integrity.

[0050] (c) Vectorization and storage: The “text2vec” model is used to convert text blocks into embeddings and store them in the “ChromaDB” vector database. The dedicated knowledge bases for different assessment agent modules need to be created and stored separately. For example, the “cognitive pattern knowledge base” corresponding to the cognitive assessment agent module mainly contains CBT-related literature, and the “behavioral performance knowledge base” corresponding to the behavioral assessment agent module mainly contains behavioral activation therapy and other content, thereby achieving precise retrieval in vertical fields. In behavioral activation therapy, the user's current behavioral state is first understood and the user is guided to complete an "Activity and Emotion Monitoring Form," such as, "To better understand your situation, can we review your main activities over the past 24 hours together? For example, what did you do from waking up to going to bed? At the same time, please rate the pleasure (P) and sense of accomplishment (M) of each activity from 0 to 10." Then, the user's avoidance patterns and triggering factors, as well as the short-term and long-term consequences of avoidance, were identified by analyzing the "Activity and Emotion Monitoring Table". High-risk situations for user avoidance, such as "turning on the computer to prepare a paper", and alternative avoidance behaviors, such as "turning on to browse the phone", were identified. Next, by systematically adding various low, medium, and high energy-consuming activities from the alternative activity library, activities that bring users pleasure and a sense of accomplishment, such as "taking a 10-minute walk," "tidying up the desk," "sending a message to a friend," and "completing a small part of an assignment," can be added to reduce avoidance behavior. The grand goals that users avoid (such as "completing the graduation thesis") are broken down into multiple small, specific, and actionable steps (such as "listing the three key points of the thesis outline"). By guiding users to rehearse obstacles, we can address potential obstacles that may arise when executing an activity plan. For example, "When you're about to 'write down three keywords,' what can we do if the thought 'What's the point of this?' pops into your head? Perhaps we can tell ourselves: 'Completing a small task is itself a victory over avoidance'?" By guiding users to complete a small task in the behavior module, the behavior assessment agent module and the cognitive assessment agent module work together to trigger positive changes in emotions and thinking, such as "Congratulations on completing 'Write down three keywords'! Think back, after you completed this small action, has your previous thought of 'I'm sure I messed up' changed in the slightest? Did this action itself challenge that thought?"

[0051] In this embodiment, step 2 also includes fine-tuning and deploying different evaluation agent modules, including steps (d) to (g): (d) Base model selection: The “Qwen-7B” base model and its tokenizer are selected as existing technologies.

[0052] (e) Command Data Construction: Each different evaluation agent module constructs an independent command fine-tuning dataset. Taking the sentiment assessment agent module as an example: Instruction: "You are a professional psychological counselor. Please analyze the following user's statement, identify their emotion type and intensity (0-10 points), and explain the reason." Type: "I have my final defense tomorrow, and my heart is pounding so hard I can't calm down at all. I'm so afraid I'll mess it up." Output: Users mentioned "final defense" (stressful event), "rapid heartbeat" (physiological reaction), and "fear of messing up" (catastrophic thinking).

[0053] (f) Fine-tuning training: LoRA (Low-Rank Adaptation) technology is used to efficiently fine-tune each different evaluation agent module to reduce computational costs and enable the derivation of multiple agent sub-modules with different professional capabilities on the same base model. LoRA (Low-Rank Adaptation) technology is an existing technology and will not be described in detail here.

[0054] (g) RAG knowledge base integration: Each fine-tuned evaluation agent module integrates its own vector database. When responding to a request, it first retrieves the top K most relevant knowledge fragments from the vector database and inputs them as context along with the user's question into the Large Language Model (LLM) to generate a professional and traceable answer. The technology of the Large Language Model (LLM) to form the answer is an existing technology.

[0055] In this embodiment, the mental health analysis and intervention system can also be connected to external expansion modules through an external system interface module. The external expansion modules include a virtual psychological companion role module, a multilingual support and international student adaptation module, and a campus mental health data dashboard module.

[0056] After connecting the virtual psychological companion role module, simulated roles with different "personalities" or "speaking styles" are generated by training a large language model (LLM), such as a "gentle counselor" or a "rational consultant", in order to increase user affinity and meet users' diverse preferences.

[0057] After connecting the multilingual support and international student adaptation module, it can serve a diverse international student population, support switching between multiple languages ​​such as English, Chinese, and Arabic, and, based on cultural sensitivity, help different users understand the differences in psychological expression and help-seeking habits in different cultural backgrounds, ensuring the accessibility and cultural appropriateness of the service.

[0058] After connecting to the campus mental health data dashboard module, it provides the university psychological counseling center management system with visualized analysis of group emotional trends, risk distribution, service usage frequency, etc., supporting macro-level decision-making and resource allocation.

[0059] In this embodiment, the multi-dimensional assessment report includes emotional state, cognitive pattern, and behavioral performance, as well as a comprehensive assessment value formed by calculating a comprehensive score S_base from the evaluation scores corresponding to emotional state, cognitive pattern, and behavioral performance. It also includes an analysis of mental health trends and integration strategies.

[0060] In one embodiment, the entire consultation process is as follows: The user wrote: "I'm going to defend my thesis tomorrow, and I'm completely unprepared. My mind is blank. I think I'm going to mess it up and not be able to graduate. My heart is pounding so hard that I can't concentrate on my presentation at all." Step 1: Parallel Evaluation and Fusion of Multiple Agents 1.1. Emotional state agent analysis.

[0061] Task: Identify and quantify emotions.

[0062] Analysis: The analysis identified "heart racing" (physiological anxiety) and "unable to calm down" (irritability).

[0063] Output: Emotional tags: ["Acute anxiety", "Panic"] Emotional intensity: 9 / 10 Basis: The user described strong physiological arousal and emotional distress.

[0064] 1.2. Cognitive Pattern Agent Analysis.

[0065] Task: Identify automatic thinking and cognitive biases.

[0066] Analysis: Typical cognitive distortions were matched from its proprietary RAG knowledge base.

[0067] “I am not prepared at all” → “All or nothing” (exaggerating the level of preparation); "I feel like I'm definitely going to mess this up" → "catastrophizing" and "negative prediction"; "Then I couldn't graduate" → "Exaggeration" (distorting the consequences of a single defense into something disastrous).

[0068] Output: Cognitive biases: ["All-or-nothing thinking", "catastrophizing", "negative prediction"]; Core belief: "I must be absolutely perfect, or it will be a complete failure."

[0069] Cognitive dimension score: 8 / 10; 1.3. Behavioral Performance Agent Analysis.

[0070] Task: Infer behavioral tendencies and functional impairments.

[0071] Analysis: "Unable to concentrate on reading the manuscript" is a typical example of behavioral avoidance and functional inhibition. This indicates that anxiety has directly impaired the final preparatory behavior.

[0072] Output: Behavioral tendency: ["Task avoidance", "Decreased efficiency"]; Specific manifestation: "Preparations for the defense were suspended"; Behavioral dimension score: 7 / 10.

[0073] 1.4. Centralized Coordination and Agent Integration Inputs: Emotion (9), Cognition (8), Behavior (7) Integration and Decision-Making: Base Score: (9*0.4) + (8*0.35) + (7*0.25) = 8.15 → High Risk.

[0074] 1.5. Rule Engine Trigger: Rule 1: Mood score > 8 → Mark as "High Acute Stress".

[0075] Rule 2: Both emotion (9) and cognition (8) are high → Identify “anxiety-cognitive vicious cycle”: catastrophic thinking exacerbates anxiety, and intense anxiety makes thinking more rigid and negative.

[0076] Rule 3: Behavior (7) indicates that "the problem is having a real impact" (preparatory work is interrupted).

[0077] Conclusion: The core driver of the current problem is "acute anxiety attacks triggered by catastrophic thinking, which have led to behavioral paralysis."

[0078] Step 2: Generation of personalized interventions based on the fusion conclusions; 2.1. Upon receiving the above fusion conclusion, the Treatment Agent retrieves and combines treatment plans from its "Intervention Responses RAG Knowledge Base": 2.1.1. Immediate Implementation (for those with highly acute emotions): Intervention: "Xiao Zhang, you sound very anxious right now, even a little panicked. This is a very real feeling when facing a major event. Let's do a simple '4-7-8' breathing exercise together to slow down your heartbeat, okay? Inhale for 4 seconds, hold your breath for 7 seconds, exhale for 8 seconds..."

[0079] Function: Quickly reduces physiological arousal levels, creating conditions for subsequent cognitive intervention. This directly targets the emotional dimension.

[0080] 2.1.2. Core Intervention (Breaking the Vicious Cycle of Cognition and Emotion): Intervention: "I've noticed your brain is telling you 'I'm definitely going to mess this up.' In psychology, this is called 'catastrophic thinking.' Can we examine this thought objectively together? For example, have you ever felt like you were going to mess up in the past, but it actually turned out okay? Or, how likely is the worst-case scenario?" Purpose: To guide cognitive restructuring and challenge automatic negative thinking. This directly targets the cognitive dimension.

[0081] 2.1.3 Behavioral Activation (Addressing Behavioral Paralysis): Intervention: "Once we let go of these exaggerated ideas a little, you can try a '5-minute start' technique. Now, just tell yourself, 'I'll look at the manuscript for 5 minutes, and after 5 minutes you are free to choose to continue or stop.' This can help us break the 'all or nothing' action pattern." Purpose: To overcome behavioral avoidance and restore its function through small tasks. This directly targets the behavioral dimension.

[0082] Step 3: Generate a comprehensive report and conclusions; The system automatically generates a structured report after the session ends.

[0083] 3.1. Psychological Health Interview Analysis Report User: Xiao Zhang (Anonymous ID: STU20241025001) Meeting time: 2024-10-24 20:15 Assessment conclusion: Acute stress state with cognitive distortion, high risk.

[0084] I. Summary of Multidimensional Assessment Overall risk index: 8.15 / 10 (high risk); Dominant problem pattern: Cognitive-emotional vicious cycle (catastrophic thinking triggers and exacerbates acute anxiety); II. Dimensional In-Depth Analysis Emotional state (9 / 10): It manifests as acute anxiety and panic, accompanied by obvious physiological symptoms (palpitations). This is the dimension that most urgently needs reassurance.

[0085] Cognitive Patterns (8 / 10): Identify catastrophic, all-or-nothing thinking. The core beliefs, centered around "perfectionism" and "fear of failure," were the main catalysts for this anxiety attack.

[0086] Behavioral performance (7 / 10): Task avoidance and behavioral inhibition (inability to prepare) are direct consequences of emotional and cognitive problems at the practical level.

[0087] III. Intervention Logic and Recommendations Immediate policy (already executed): Objective: To reduce emotional arousal.

[0088] Measures: 4-7-8 breathing technique, mindfulness grounding techniques.

[0089] Core strategy (recommended to continue): Objective: To correct cognitive biases.

[0090] Measures: Guide students to practice using a "mind log" to learn to identify and challenge catastrophic thinking.

[0091] Support strategy (recommended to continue): Objective: To restore social functioning.

[0092] Measures: Implement "behavioral activation" and "task decomposition", breaking down "preparing for the defense" into smaller steps such as "organizing the outline" and "practicing the opening 5 minutes".

[0093] IV. Referral and Early Warning Given the current high-risk situation, the system has initiated enhanced monitoring.

[0094] Strongly recommended: After your thesis defense, schedule an in-person consultation at the university's counseling center to address your underlying perfectionist tendencies and stress coping mechanisms.

[0095] The working principle of this invention is as follows: First, the initial triage agent module triages the user's input information and dialogue content. Then, multiple different evaluation agent modules perform parallel and independent multi-dimensional psychological assessments, evaluating the user's current situation from three aspects: emotion, cognition, and behavior. These assessment values ​​are then weighted and summed with preset weights to obtain a comprehensive structured result. An intervention plan is selected based on this comprehensive structured result and the three assessment values. This approach combines the comprehensive structured result with intuitive judgment based on the three assessment values. When a problem occurs with a significant emotion assessment value, an emotion adjustment intervention plan is directly initiated. Similarly, when cognitive and behavioral assessment values ​​are high, cognitive and behavioral interventions are also considered. After the intervention plan is implemented, the emotion, cognition, and behavioral assessment values ​​are re-evaluated, and adjustments to the short-term and long-term dialogue intervention plans are determined based on these values. This ensures that the intervention plan can adapt to a normal state after adjustment, preventing over-intervention that could lead to inaccurate results.

Claims

1. A method for mental health analysis and intervention based on multi-agent collaborative assessment, implemented through a mental health analysis and intervention system, characterized by: Includes the following steps: S1 acquires the user identity information and corresponding input information of the user entering the mental health analysis and intervention system, and performs preprocessing through the initial triage Agent module; S2 evaluates user questions based on behavior, cognition, and emotion after initial triage, obtaining corresponding evaluation values. It then performs in-depth analysis on the vertical domain-specific RAG knowledge base that matches the behavior, cognition, and emotion, and sums the results with the corresponding evaluation values ​​according to the preset weights of behavior, cognition, and emotion to obtain a comprehensive structured evaluation result. The S3 central coordination and treatment agent receives the comprehensive structured assessment results from step S2, as well as the assessments of behavior, cognition, and emotion, to obtain corresponding assessment values. Based on the assessment results and the corresponding assessment values ​​of behavior, cognition, and emotion, the agent recommends intervention priorities in the multi-dimensional assessment report corresponding to the generated user identity information. S4 uses the treatment intervention agent module to retrieve data from the intervention measures RAG knowledge base and form a matching intervention strategy, and then forms a corresponding intervention plan; S5 dynamically adjusts the weights corresponding to behavior, cognition, and emotion based on the post-intervention situation, and performs adaptive intervention for users in short-term and long-term dialogue interactions.

2. The mental health analysis and intervention method based on multi-agent collaborative assessment according to claim 1, characterized in that: In step S1, the user connects to an external identity authentication system through an identity authentication interface to securely and independently log in to the mental health analysis and intervention system. In step S1, after the user logs into the mental health analysis and intervention system, and with the user's authorization and in compliance with the security and privacy protection module, the mental health analysis and intervention system exchanges non-anonymized information with the mental health counseling center management system that matches the external identity authentication system. After preprocessing by the initial triage Agent module in step S1, the process also includes: performing initial triage for intent identification; if the user is assessed as being at high risk, the crisis protocol of the mental health analysis and intervention system is triggered; the mental health analysis and intervention system sends encrypted alarm information to the mental health counseling center management system or the user's emergency contact via API call, SMS, or email gateway. Otherwise proceed to step S2.

3. The mental health analysis and intervention method based on multi-agent collaborative assessment according to claim 1, characterized in that: Step S1 further includes anonymizing the user-input data through a security and privacy protection module: encrypting data transmission and static data, and after the user completes the dialogue, only the desensitized structured assessment data is retained in the mental health analysis and intervention system for long-term trend analysis, while the non-desensitized structured assessment data is automatically deleted.

4. The mental health analysis and intervention method based on multi-agent collaborative assessment according to claim 1, characterized in that: Step S2 further includes: steps S2.1 to S2.

3. When a user is assessed as not at high risk, S2.1 engages in dialogue with the user through a Large Language Model (LLM) and generates an initial response. Then, based on a vertical domain-specific RAG knowledge base matched with different assessment agent modules, it independently retrieves the data information from the user's response and generates structured assessment results for the corresponding behavioral, cognitive, and emotional dimensions. S2.2 Based on the evaluation values ​​from the structured evaluation results of three different dimensions, three different corresponding weights W_b, W_c, and W_e are matched, and the sum of the three different weights is 1. Then, the comprehensive evaluation value S_base is calculated using formula (1). S_base = (Score_e * W_e) + (Score_c * W_c) + (Score_b * W_b) (1), Score_e represents the evaluation value of the structured assessment result in the emotion dimension, and W_e represents the weight corresponding to the emotion dimension, with a value range of 0.3 - 0.4; Score_c represents the evaluation value of the structured assessment result in the cognition dimension, and W_c represents the weight corresponding to the cognition dimension, with a value range of 0.3 - 0.4; Score_b represents the evaluation value of the structured assessment result in the behavior dimension, and W_b represents the weight corresponding to the behavior dimension, with a value range of 0.3 - 0.

4. S2.3 By using a Large Language Model (LLM) to integrate structured assessment results, personalized responses that match the user are generated, and corresponding personalized intervention plans are provided to the user.

5. The mental health analysis and intervention method based on multi-agent collaborative assessment according to claim 4, characterized in that: Step S2.2 further includes: steps S2.2.1 to S2.2.

3. S2.2.1 If the evaluation value of any dimension in the three different dimensions of the structured evaluation results is > 9, the comprehensive evaluation value S_base becomes invalid and the crisis protocol is triggered; S2.2.2 If the structured assessment result in the emotion dimension is > 7 and the structured assessment result in the cognition dimension is > 7, the resulting comprehensive evaluation result is a high-risk level, and a cognitive behavioral therapy (CBT) intervention plan is provided to the user in the resulting assessment report. S2.2.3 If the structured assessment result in the emotion dimension is > 7 and the structured assessment result in the behavior dimension is < 4, then the user will be provided with a behavioral activation intervention plan in the generated assessment report.

6. The mental health analysis and intervention method based on multi-agent collaborative assessment according to claim 1, characterized in that: Step S3 further includes: The psychological profile is dynamically updated in real time based on the dynamic data during the initial triage process of the Agent module in step S1, the process of generating personalized responses for users by integrating the large language model (LLM) in step S2, and the process of users selecting the appropriate personalized intervention path.

7. The mental health analysis and intervention method based on multi-agent collaborative assessment according to claim 1, characterized in that: The multi-dimensional assessment report in step S3 includes emotional state, cognitive pattern, and behavioral performance, as well as a comprehensive assessment value S_base calculated from the assessment values ​​corresponding to emotional state, cognitive pattern, and behavioral performance.

8. The method for mental health analysis and intervention based on multi-agent collaborative assessment according to claim 1, characterized in that: Step S4 includes S4.1 to S4.

3. The S4.1 Treatment Intervention Agent module retrieves corresponding emotional states, cognitive patterns, and behavioral performances from the intervention measures RAG knowledge base based on the multi-dimensional assessment report, and recommends corresponding intervention strategies through immediate intervention goals corresponding to emotional states, core intervention goals corresponding to cognitive patterns, and supportive intervention goals corresponding to behavioral performances. S4.2 In the cognitive pattern retrieval process, the user's verbal responses are mapped to standardized cognitive distortion types in cognitive behavioral therapy (CBT) to obtain different degrees of cognitive bias. S4.3 Based on different degrees of cognitive bias, corresponding intervention plans are formed, and personalized intervention plan outputs are formed by combining the intervention plan corresponding to the immediate intervention goal and the intervention plan corresponding to the supporting intervention goal.

9. A method for mental health analysis and intervention based on multi-agent collaborative assessment according to claim 8, characterized in that: Step S4.2 includes steps S4.2.1 to S4.2.

3. S 4.2.1 The cognitive assessment agent module performs word segmentation and syntactic analysis on the user's input language response to extract the core viewpoints, beliefs or predictions expressed by the user; S 4.2.2 will extract the core viewpoints, beliefs or predictions expressed by users and search and match them with the RAG knowledge base of specific intervention measures; S 4.2.3 Compare the semantic vectors of the user’s core viewpoints, beliefs or predictions with the semantic vectors of example patterns in the RAG knowledge base of the specific intervention measures, and then use the numerical value of the similarity to judge the different degrees of cognitive bias.

10. A method for mental health analysis and intervention based on multi-agent collaborative assessment according to claim 1, characterized in that: Step S5 also includes steps S5.1 to S5.

3. During short-term dialogue interactions, S5.1 users are evaluated through the Emotion Assessment Agent, Cognitive Assessment Agent, and Behavioral Assessment Agent modules, and corresponding weights are assigned. After the crisis intervention mode is triggered, the corresponding weight ratio is reduced so that the weight ratio corresponding to the emotion assessment value is equal to the sum of the weight ratio corresponding to the cognitive assessment value and the weight ratio corresponding to the behavioral assessment value, and the user is guided to implement the corresponding intervention plan. After intervention, a reassessment is conducted until the emotional crisis is resolved. The weighting ratio is then readjusted, with cognition-behavior as the dominant factor, to guide users in cognitive restructuring and behavioral planning. S5.2 For users with long-term dialogue interactions, the weights corresponding to the behaviors, cognitions, and emotions that have the best effects in historical data are adjusted. If it is detected that the evaluation value of different adaptive strategies decreases in subsequent dialogue interactions, the weights corresponding to the original behaviors, cognitions, and emotions are adjusted back.

Images