Progressive stress resistance evaluation system based on large model agent

Abstract

This invention provides a progressive stress resilience assessment system based on a large-scale intelligent agent; it primarily addresses the problems of limited interaction methods and lack of dynamic adaptability in existing psychological state assessments. The system employs a digital human interactive agent, an emotion perception and scoring agent, an emotion guidance and planning agent, an empathy expression generation agent, and a scale scoring agent. The emotion perception and scoring agent calculates negative emotion scores based on questions and answers. The emotion guidance and planning agent generates the next question based on semantic understanding and score feedback, achieving an adaptive assessment path. The empathy expression generation agent generates system responses with emotional connection and guidance based on multiple inputs from the current question, user answers, and the next question. The digital human interactive agent presents outputs via voice and facial expressions, achieving natural interaction. The scale scoring agent manages and scores the scale items for the ruminative thinking and stress resilience stages, ultimately obtaining the stress resilience assessment result.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a progressive stress resistance assessment system based on a large model intelligent agent. Background Technology

[0002] With the accelerating pace of society and increasing external pressures, the demand for psychological state assessment is constantly growing in fields such as education, corporate management, and human-computer interaction. Existing methods mostly rely on fixed questionnaires or single interviews. Although they are simple to operate, they suffer from problems such as limited interaction, difficulty in dynamically adapting to individual differences, lack of multimodal perception, and rigid feedback methods, making it difficult to comprehensively reflect an individual's true psychological state.

[0003] In the field of artificial intelligence, an agent is an autonomous unit capable of perceiving its environment, making decisions based on input, and invoking tools or performing actions. Compared to traditional static programs, agents possess stronger context awareness and task adaptability, and can flexibly adjust strategies in complex dialogues and interactions. Collaboration among multiple agents can also form task division and feedback loops, making it possible to build complex evaluation systems.

[0004] The LangGraph framework proposed by the LangChain team supports process-oriented, phased task execution and persistent state management through state graph management and multi-agent collaboration mechanisms. This provides a new approach to psychological state assessment, enabling the system to move beyond the "single question-and-answer - static scoring" model and towards phased, adaptive, and multimodal dynamic assessment.

[0005] However, current agent-based psychological state assessment applications still have shortcomings: they lack a systematic multi-tool collaborative architecture, have limited context-aware decision-making, and fail to form a closed loop between multimodal interaction and digital human interfaces, resulting in a still somewhat rigid user experience. Therefore, it is necessary to design an agent-based negative emotion and stress resilience assessment system, using a phased strategy driven by LangGraph and multimodal interaction to improve the scientific rigor, robustness, and user acceptance of the assessment. Summary of the Invention

[0006] To overcome the problems of limited interaction methods and lack of dynamic adaptability in existing psychological state assessments, this invention proposes a progressive stress resilience assessment system based on large-scale intelligent agents. This system relies on the LangGraph framework and employs multi-agent collaboration and phased assessment as its core strategies to construct a progressive assessment chain from negative emotions to stress resilience.

[0007] The progressive stress resilience assessment system based on a large-scale model agent includes a digital human interaction agent, an emotion perception and scoring agent, an emotion guidance and planning agent, an empathy expression generation agent, and a scale scoring agent. The emotion perception and scoring agent calculates negative emotion scores based on questions and answers. The emotion guidance and planning agent generates the next question based on semantic understanding and score feedback, achieving an adaptive assessment path. The empathy expression generation agent generates system responses with emotional connection and guidance based on multiple inputs from the current question, user answers, and the next question. The digital human interaction agent presents output through voice and facial expressions, achieving natural interaction. The scale scoring agent manages and scores the scale items for the ruminative thinking and stress resilience stages.

[0008] The progressive stress resilience assessment system based on large-scale intelligent agents maintains the current stage, problem pointers, and scores for each index through the LangGraph state mechanism, enabling traceability and dynamic control of the task flow. When the total negative emotion score exceeds a set threshold, the system automatically enters the ruminant thinking detection stage; otherwise, it directly proceeds to the stress resilience assessment. Through this state-driven closed-loop design, this invention significantly improves the naturalness and interactive experience of the assessment process while maintaining scientific rigor and robustness.

[0009] The technical solution adopted by this invention to solve its technical problem includes the following steps:

[0010] Step S1: Construct the dataset;

[0011] We collected the following raw data: the voice responses of visitors during consultations, the PHQ-9 scale scores completed by visitors, and the consultation dialogue text after compliance-compliant anonymization. We then used Whisper to transcribe the voice responses from the raw data into text, removing samples with transcription failures and missing values. Based on the PHQ-9 scale scores, we mapped user responses to hierarchical labels. We then used the hierarchical labels, question text, response text, and topic category as a set of data. Based on these sets of data, we constructed a training set.

[0012] Step S2: Construct a multi-agent collaborative network; the multi-agent collaborative network is constructed based on the LangGraph framework; the multi-agent collaborative network is a state machine; the multi-agent collaborative system includes several nodes and edges between nodes; nodes realize data flow and state transfer through edges; the nodes of the multi-agent collaborative network include a digital human interaction agent, an emotion perception and scoring agent, an emotion guidance and planning agent, an empathy expression generation agent, a scale scoring agent, a state object, and a question list; train the emotion perception and scoring agent using a dataset, and obtain the optimal emotion perception and scoring agent after reaching a predetermined training threshold; replace the emotion perception and scoring agent in the multi-agent collaborative network with the optimal emotion perception and scoring agent;

[0013] Step S3: Use a multi-agent collaborative network to detect the user's stress resistance and obtain the detection results.

[0014] Furthermore, the digital human interactive agent is constructed based on the ER-NeRF model; the input of the digital human interactive agent is speech text; the digital human interactive agent outputs natural speech and facial expressions through TTS speech synthesis method, which are used to present questions, empathic feedback and scale items to users;

[0015] The emotion perception and scoring agent is based on the vertical domain large model ChatGLM4-9B, which is efficiently fine-tuned with LoRA+ parameters. The input of the emotion perception and scoring agent is the question text and the user's answer text. The output of the emotion perception and scoring agent is the negative emotion type and the negative emotion score. The negative emotions include the negative emotion index, the anhedonia index, the physical distress index, and the high-risk behavior index. The negative emotion score is in the range of [1, 4]. The emotion perception and scoring agent writes the negative emotion score into the state object according to the negative emotion type using the state update method.

[0016] The negative emotion index S_NE (Negative Emotion Sub-Index) reflects the intensity of expressions related to low mood; the anhedonia index S_ANH (Anhedonia Sub-Index) reflects the degree of reduced interest / pleasure in daily activities; the somatic distress index S_SOM (Somatic Distress Sub-Index) reflects the intensity of physiological discomfort related to mental and physical stress; and the risk behavior index S_RISK (Risk Behavior Sub-Index) reflects the intensity of potential high-risk behaviors such as impulsivity / risk-taking.

[0017] The emotion guidance planning agent is based on DeepSeek-R1; the input of the emotion guidance planning agent is the current question, the user's answer text, and the negative emotion index, anhedonia index, physical distress index, and high-risk behavior index; the output of the emotion guidance planning agent is the question text.

[0018] The emotion-guiding planning agent checks whether all questions in the current index category of the question list have been asked. If all questions have been asked, it moves to the next index category; if not, it extracts the next question from the index category of the question list. It iterates through all index categories of questions until all questions in the question list have been asked.

[0019] Specifically, in the Physical Distress Index category, if the first question is "How is your physical condition recently, and does it affect your life?", and the score for the first question is 1, it means that the user has not experienced any obvious physical discomfort, and the user is moved to the next index category; if the score for the first question is > 1, then the other questions in the index category are traversed.

[0020] If the current index is negative sentiment, the next index category will be anhedonia index.

[0021] If the current index is the anhedonia index, calculate the sum of the user's scores on the negative emotion index and the anhedonia index. If the sum of the scores on the negative emotion index and the anhedonia index is less than or equal to 2, it is determined that the current stage does not meet the screening criteria for depressive episodes; if the sum of the scores on the negative emotion index and the anhedonia index is greater than 2, the next index category is the physical distress index.

[0022] The empathy expression generation agent is based on DeepSeek-R1; the input of the empathy expression generation agent is the current question text, the user's answer text, and the next question text; the output of the empathy expression generation agent is the natural response text, which is then used as speech text and input into the digital human interaction agent.

[0023] The scale scoring and management agent includes a ruminant index. Calculation module and stress resistance index Calculation module; Rumination index The calculation module calculates the Rumination Index based on the scores of questions in the Rumination RRS-10 scale. Compression resistance index The calculation module calculates the resilience index based on the scores of questions in the simplified psychological resilience scale CD-RISC-10. ; Rumination Index With the state object Perform summation, and use the result of the summation as the new... Write to the state object; set the resilience index With the state object Perform summation, and use the result of the summation as the new... Write the state object; the input for the scale rating and management agent is the negative emotion type and negative emotion score;

[0024] The state object is:

[0025] };in, The index categories represent the current issues; these categories include negative emotion index, anhedonia index, physical distress index, high-risk behavior index, rumination index, and stress resilience index. This represents the negative sentiment index value. This represents the state value of the anhedonia index. The physical distress index is the state value. This represents the risk index status value for high-risk behaviors. The total score represents negative emotions. This represents the rumination index state value. This represents the compressive strength index state value.

[0026] The state update method is as follows:

[0027] If the negative emotion type is a negative emotion index, then accumulate the values ​​in the state object. The value will be added to the state object. The value is divided by the total number of questions under the negative sentiment index question category;

[0028] If the negative emotion type is the anhedonia index, then the index is accumulated in the state object. The value will be added to the state object. The value is divided by the total number of problems under the Anhedonia Index problem category;

[0029] If the negative emotion type is the physical distress index, then accumulate the values ​​in the state object. The value will be added to the state object. The value is divided by the total number of problems under the Physical Distress Index problem category;

[0030] If the negative emotion type is a high-risk behavior risk index, then the index is accumulated in the state object. The value will be added to the state object. The value is divided by the total number of questions under the high-risk behavior risk index question category;

[0031] The list of questions is based on the MINI scale; the MINI scale is the Mini-International Neuropsychiatric Interview scale.

[0032] The issues in the three items of depressive episode, dysphoric mood and suicide in the MINI scale were reclassified into four dimensions according to the negative emotion index, anhedonia index, physical distress index and high-risk behavior risk index.

[0033] The negative emotion index dimension includes questions such as "How have you been feeling lately, and has it affected your life?" and "Have you felt depressed or down for most of the day in the past two weeks?"

[0034] Questions for the anhedonia index dimension include "Do you usually enjoy doing things, and have you done any of these things recently?" and "What types of TV programs do you usually like to watch?"

[0035] Questions in the Physical Distress Index dimension include: "How is your physical condition lately, and has it affected your life?", "What do you want to do when you can't fall asleep?", "Have you had a poor appetite or overeaten lately?", "Do you feel tired and lack energy every day?", "Have you recently had difficulty concentrating, hesitated, or had difficulty making decisions?"

[0036] The high-risk behavior risk index includes questions such as "What are your views on suicide?" and "Have you recently had recurring thoughts of self-harm, suicide, or wishing to die?"

[0037] By default, multi-agent cooperative networks start with questions about "negative emotions".

[0038] Furthermore, the steps for using a multi-agent cooperative network to detect a user's resilience are as follows:

[0039] Step S3-1: Detect the negative emotion stage;

[0040] Step S3-1-1: The digital human interactive intelligent agent outputs the first question;

[0041] The emotion-guided planning agent selects question text from the question list; if it is the first time asking a question, it uses a question related to "negative emotions" as the opening question to obtain the user's voice response.

[0042] If this is not the first time a question has been asked, retrieve it from the state object. Proceed to step S3-1-4 to obtain the next question text;

[0043] Step S3-1-2: Based on the user's voice response, use the Whisper speech recognition model to transcribe the user's voice response into speech text;

[0044] Step S3-1-3: Analysis and quantification of the agent for emotion perception scoring;

[0045] Input the voice text and the corresponding question text into the emotion perception and scoring agent; the emotion perception and scoring agent outputs a negative emotion score based on the type of negative emotion.

[0046] Step S3-1-4: The emotion-guided planning agent generates the next question;

[0047] Input the negative emotion score into the emotion-guided planning agent; the emotion-guided planning agent outputs the text of the next question;

[0048] Step S3-1-5: Express empathy, generate an intelligent agent to optimize the response, and provide feedback to the user;

[0049] Input the text of the next question in S3-1-4 into the empathic expression to generate an agent and obtain a natural response text;

[0050] The logic for generating natural response text includes an emotional feedback layer and a guidance layer. The emotional feedback layer responds to the user's answer, demonstrating understanding and emotional support. The guidance layer embeds a transitional phrase for the next question at the end of the response to maintain contextual coherence and emotional rhythm. The natural response text is input into the digital human interaction agent for speech synthesis and facial expression driving, achieving a unified semantic and visual empathetic communication experience through streaming video. If the natural response text does not contain the next question, proceed to step 2; if the natural response text does contain the next question, proceed to step 1-1.

[0051] Step S3-2: Rumination detection phase;

[0052] The total negative affect score is calculated as the Negative Affect Total Score. :

[0053]

[0054] in The index score for the k-th topic;

[0055] When the total score for negative emotions If the score is less than the preset threshold of 8 points, proceed to step 3;

[0056] When the total score for negative emotions When the score exceeds a preset threshold of 8 points, the digital human interaction agent presents the user with RRS-10 questions. The user answers by clicking on the screen, avoiding the additional cognitive load of voice input; ruminative thinking index The calculation module calculates the ruminant thinking index based on the user's clicks and answers. ;

[0057] Step S3-3: Compression resistance testing stage;

[0058] The digital human interactive agent presents users with questions from the CD-RISC-10 scale, and users respond using a screen-clicking mode; stress resilience index. The calculation module calculates the stress resistance index based on the user's clicks and answers. ;

[0059] Step S3-4: Summarize the results;

[0060] If the stress resistance index When the value is greater than or equal to 40, the user has a strong ability to withstand pressure.

[0061] If the stress resistance index If the value is greater than or equal to 30 and less than 40, then the user's stress resistance is within the normal range.

[0062] If the stress resistance index If the value is less than 30, the user's stress tolerance is relatively weak.

[0063] Furthermore, the emotion perception and scoring agent is based on the vertical domain large model ChatGLM4-9B, which is efficiently fine-tuned using LoRA+ parameters; the fine-tuning process of the emotion perception and scoring agent is as follows:

[0064] The parameter matrix of the large vertical domain model ChatGLM4-9B is decomposed using LoRA+:

[0065] The parameter matrix is , The update format is as follows:

[0066]

[0067] in, These are the frozen pre-training parameters. This indicates a low-rank approximate update;

[0068] Use LoRA+ to The solution is:

[0069] , , , ≪ ;

[0070] in and All are low-rank decomposition matrices; for The rank of the matrix after low-rank decomposition; for The dimension;

[0071] Original calculation intermediate state for:

[0072]

[0073] For the input vector 𝑥, during forward propagation, the original intermediate states are computed. Updated to:

[0074]

[0075] After training, the original parameter matrix and the weights obtained during training and The updated parameter matrix W is calculated as follows:

[0076] ;

[0077] For standard LoRA, the same learning rate for A and B has been shown to lead to suboptimal learning when the embedding dimension is large. In LoRA+, the learning rates for A and B are set as follows:

[0078]

[0079] Let A be the learning rate; Let B be the learning rate; A constant much greater than 1; this mechanism can alleviate the convergence problem under high-dimensional embedding and improve fine-tuning efficiency.

[0080] The beneficial effects of this invention are as follows: By employing the multi-agent orchestration mechanism and phased dynamic evaluation strategy of the LangGraph framework, this invention achieves structured, traceable, and adaptive decision-making in the psychological state assessment process. By constructing a multi-agent collaborative network including an emotion perception and scoring agent, an emotion guidance and planning agent, an empathy expression generation agent, a digital human interaction agent, and a scale scoring and management agent, different functional modules achieve data flow closure and phased transitions within a unified state graph, thereby overcoming the problems of "fixed question-and-answer process, lack of semantic feedback, and insufficient response to individual differences" in traditional psychological assessments. Furthermore, it introduces LoRA+-based vertical domain model fine-tuning to ensure efficient generalization and emotion recognition capabilities under limited computing power; combined with ER-NeRF digital human interaction technology and a speech synthesis interface, it achieves natural, continuous speech and expression-driven human-computer communication. Attached Figure Description

[0081] Figure 1 This is a diagram illustrating the overall structure of the method of the present invention;

[0082] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0083] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0084] The progressive stress resistance assessment system based on large model agents includes a multi-agent collaborative network; the multi-agent collaborative network is built on the LangGraph framework; the multi-agent collaborative network is a state machine; the multi-agent collaborative system includes several nodes and edges between nodes; nodes realize data flow and state transmission through edges.

[0085] The nodes of the multi-agent collaborative network include digital human interaction agent, emotion perception and scoring agent, emotion guidance and planning agent, empathy expression generation agent, scale scoring agent, state object and question list;

[0086] The digital human interactive agent is constructed based on the ER-NeRF model; the input of the digital human interactive agent is speech text; the digital human interactive agent outputs natural speech and facial expressions through TTS speech synthesis method, which is used to present questions, empathic feedback and scale items to users;

[0087] The emotion perception and scoring agent is based on the vertical domain large model ChatGLM4-9B, which is efficiently fine-tuned with LoRA+ parameters. The input of the emotion perception and scoring agent is the question text and the user's answer text. The output of the emotion perception and scoring agent is the negative emotion type and the negative emotion score. The negative emotions include the negative emotion index, the anhedonia index, the physical distress index, and the high-risk behavior index. The negative emotion score is in the range of [1, 4]. The emotion perception and scoring agent writes the negative emotion score into the state object according to the negative emotion type using the state update method.

[0088] The negative emotion index S_NE (Negative Emotion Sub-Index) reflects the intensity of expressions related to low mood; the anhedonia index S_ANH (Anhedonia Sub-Index) reflects the degree of reduced interest / pleasure in daily activities; the somatic distress index S_SOM (Somatic Distress Sub-Index) reflects the intensity of physiological discomfort related to mental and physical stress; and the risk behavior index S_RISK (Risk Behavior Sub-Index) reflects the intensity of potential high-risk behaviors such as impulsivity / risk-taking.

[0089] The emotion guidance planning agent is based on DeepSeek-R1; the input of the emotion guidance planning agent is the current question, the user's answer text, and the negative emotion index, anhedonia index, physical distress index, and high-risk behavior index; the output of the emotion guidance planning agent is the question text.

[0090] The emotion-guiding planning agent checks whether all questions in the current index category of the question list have been asked. If all questions have been asked, it moves to the next index category; if not, it extracts the next question from the index category of the question list. It iterates through all index categories of questions until all questions in the question list have been asked.

[0091] Specifically, in the Physical Distress Index category, if the first question is "How is your physical condition recently, and does it affect your life?", and the score for the first question is 1, it means that the user has not experienced any obvious physical discomfort, and the user is moved to the next index category; if the score for the first question is > 1, then the other questions in the index category are traversed.

[0092] If the current index is negative sentiment, the next index category will be anhedonia index.

[0093] If the current index is the anhedonia index, calculate the sum of the user's scores on the negative emotion index and the anhedonia index. If the sum of the scores on the negative emotion index and the anhedonia index is less than or equal to 2, it is determined that the current stage does not meet the screening criteria for depressive episodes; if the sum of the scores on the negative emotion index and the anhedonia index is greater than 2, the next index category is the physical distress index.

[0094] The empathy expression generation agent is based on DeepSeek-R1; the input of the empathy expression generation agent is the current question text, the user's answer text, and the next question text; the output of the empathy expression generation agent is the natural response text, which is then used as speech text and input into the digital human interaction agent.

[0095] The scale scoring and management agent includes a ruminant index. Calculation module and stress resistance index Calculation module; Rumination index The calculation module calculates the Rumination Index based on the scores of questions in the Rumination RRS-10 scale. Compression resistance index The calculation module calculates the resilience index based on the scores of questions in the simplified psychological resilience scale CD-RISC-10. ; Rumination Index With the state object Perform summation, and use the result of the summation as the new... Write to the state object; set the resilience index With the state object Perform summation, and use the result of the summation as the new... Write the state object; the input for the scale rating and management agent is the negative emotion type and negative emotion score;

[0096] The state object is:

[0097] };in, The index categories represent the current issues; these categories include negative emotion index, anhedonia index, physical distress index, high-risk behavior index, rumination index, and stress resilience index. This represents the negative sentiment index value. This represents the state value of the anhedonia index. The physical distress index is the state value. This represents the risk index status value for high-risk behaviors. The total score represents negative emotions. This represents the rumination index state value. This represents the compressive strength index state value.

[0098] The state update method is as follows:

[0099] If the negative emotion type is a negative emotion index, then accumulate the values ​​in the state object. The value will be added to the state object. The value is divided by the total number of questions under the negative sentiment index question category;

[0100] If the negative emotion type is the anhedonia index, then the index is accumulated in the state object. The value will be added to the state object. The value is divided by the total number of problems under the Anhedonia Index problem category;

[0101] If the negative emotion type is the physical distress index, then accumulate the values ​​in the state object. The value will be added to the state object. The value is divided by the total number of problems under the Physical Distress Index problem category;

[0102] If the negative emotion type is a high-risk behavior risk index, then the index is accumulated in the state object. The value will be added to the state object. The value is divided by the total number of questions under the high-risk behavior risk index question category;

[0103] The list of questions is based on the MINI scale; the MINI scale is the Mini-International Neuropsychiatric Interview scale.

[0104] The MINI scale includes four dimensions: negative emotion index, anhedonia index, physical distress index, and high-risk behavior risk index.

[0105] The negative sentiment index includes questions such as "How have you been feeling lately, and has it affected your life?" and "Have you felt depressed for most of the past two weeks?".

[0106] Questions for the anhedonia index dimension include "Do you usually enjoy doing things? Have you done any of these things recently?";

[0107] The Physical Distress Index dimension includes questions such as "Have you recently experienced sleep or fatigue problems?";

[0108] The high-risk behavior risk index includes questions such as "When under increased stress, have you ever had thoughts of harming yourself or others?"

[0109] By default, multi-agent cooperative networks begin with questions related to "negative emotions."

[0110] The method for assessing resilience using multi-agent cooperative networks includes the following steps:

[0111] Step 1: Detect the negative emotion stage;

[0112] Step 1-1: The digital human interactive intelligent agent outputs the first question;

[0113] The emotion-guided planning agent selects question text from the question list; if it is the first time asking a question, it uses a question related to "negative emotions" as the opening question to obtain the user's voice response.

[0114] If this is not the first time a question has been asked, retrieve it from the state object. Jump to steps 1-4 to obtain the next question text;

[0115] Steps 1-2: Based on the user's voice response, use the Whisper speech recognition model to transcribe the user's voice response into speech text;

[0116] Steps 1-3: Analysis and quantification of the emotion perception scoring agent;

[0117] Input the voice text and the corresponding question text into the emotion perception and scoring agent; the emotion perception and scoring agent outputs a negative emotion score based on the type of negative emotion.

[0118] Steps 1-4: The emotion-guided planning agent generates the next question;

[0119] Input the negative emotion score into the emotion-guided planning agent; the emotion-guided planning agent outputs the text of the next question;

[0120] Steps 1-5: Empathic expression generates an intelligent agent that optimizes responses and provides feedback to the user;

[0121] Input the text of the next question from 1-4 into the empathy expression to generate an agent and obtain a natural response text;

[0122] The logic for generating natural response text includes an emotional feedback layer and a guidance layer. The emotional feedback layer responds to the user's answer, demonstrating understanding and emotional support. The guidance layer embeds a transitional phrase for the next question at the end of the response to maintain contextual coherence and emotional rhythm. The natural response text is input into the digital human interaction agent for speech synthesis and facial expression driving, achieving a unified semantic and visual empathetic communication experience through streaming video. If the natural response text does not contain the next question, proceed to step 2; if the natural response text does contain the next question, proceed to step 1-1.

[0123] Step 2: Rumination Test Phase;

[0124] The total negative affect score is calculated as the Negative Affect Total Score. :

[0125]

[0126] in The index score for the k-th topic;

[0127] When the total score for negative emotions If the score is less than the preset threshold of 8 points, proceed to step 3;

[0128] When the total score for negative emotions When the score exceeds a preset threshold of 8 points, the digital human interaction agent presents the user with RRS-10 questions. The user answers by clicking on the screen, avoiding the additional cognitive load of voice input; ruminative thinking index The calculation module calculates the ruminant thinking index based on the user's clicks and answers. ;

[0129] By introducing a rumination test after negative emotion assessment, the system achieves a progressive evaluation from emotional experience to cognitive processing, revealing the indirect mechanism by which negative emotions lead to functional impairment, and further enhancing the scientific rigor and robustness of the overall assessment system. The design of the rumination test stage is based on the chain mechanism in psychological research that "elevated levels of negative emotions often lead individuals to repeatedly think about negative events, thereby weakening adaptive function," aiming to identify potential cognitive solidification risks caused by negative emotions.

[0130] The rumination assessment phase used the RRS-10 (Ruminative Responses Scale-10). The RRS-10 is a simplified version of the original 22-item rumination response scale, retaining 10 highly representative items to distinguish between two types of rumination components:

[0131] Reflective thinking: It reflects an individual's attempt to understand or solve problems through self-analysis and has a certain adaptive function;

[0132] Brooding: Characterized by negative self-comparison and recurring brooding, it is usually associated with impaired psychological functioning;

[0133] Step 3: Compression resistance testing stage;

[0134] The digital human interactive agent presents users with questions from the CD-RISC-10 scale, and users respond using a screen-clicking mode; stress resilience index. The calculation module calculates the stress resistance index based on the user's clicks and answers. ;

[0135] The stress resistance assessment phase aims to evaluate an individual's psychological resilience and adaptability when facing stress, setbacks, or unexpected events, serving as a positive control indicator in the overall assessment system.

[0136] The stress resilience assessment phase used the Connor-Davidson Resilience Scale-10 (CD-RISC-10) as the core tool. The CD-RISC-10 was revised by Connor and Davidson based on the original 25-item scale, retaining 10 core items and employing a five-point Likert scale (1 = never, 5 = always), covering four key dimensions: self-efficacy; emotional control; adaptability; and tenacity.

[0137] Step 4: Summarize the results;

[0138] If the stress resistance index When the value is greater than or equal to 40, the user has a strong ability to withstand pressure.

[0139] If the stress resistance index If the value is greater than or equal to 30 and less than 40, then the user's stress resistance is within the normal range.

[0140] If the stress resistance index If the score is less than 30, the user's stress resistance is relatively weak;

[0141] By incorporating negative emotion index, rumination index, and stress resilience index into a multi-agent system, the system can not only quantify an individual's emotional vulnerability but also capture their resilience potential in a balanced way. This results in a comparative analysis showing that "high negative emotion leads to strong rumination and consequently low stress resilience," and "low negative emotion leads to high stress resilience," revealing a chain-like psychological regulation mechanism. The system displays the sub-results and overall trends through a visual interface, providing feedback in chart or text format. This information is for psychological state reference only and does not have medical diagnostic value. This stage achieves a closed loop from multi-stage assessment to result output, ensuring the integrity and interpretability of the assessment process.

[0142] Building the training set:

[0143] We collected the following raw data: the voice responses of visitors during consultations, the PHQ-9 scale scores completed by visitors, and the consultation dialogue text after compliance-compliant anonymization. We then used Whisper to transcribe the voice responses from the raw data into text, removing samples with transcription failures and missing values. Based on the PHQ-9 scale scores, we mapped user responses to hierarchical labels. We then used the hierarchical labels, question text, response text, and topic category as a set of data. Based on these sets of data, we constructed a training set.

[0144] The emotion perception and scoring agent is based on the vertical domain large model ChatGLM4-9B, which is efficiently fine-tuned using LoRA+ parameters; the fine-tuning process of the emotion perception and scoring agent is as follows:

[0145] The parameter matrix of the large vertical domain model ChatGLM4-9B is decomposed using LoRA+:

[0146] The parameter matrix is , The update format is as follows:

[0147]

[0148] in, These are the frozen pre-training parameters. This indicates a low-rank approximate update;

[0149] Use LoRA+ to The solution is:

[0150] , , , ≪ ;

[0151] in and All are low-rank decomposition matrices; for The rank of the matrix after low-rank decomposition; for The dimension;

[0152] Original calculation intermediate state for:

[0153]

[0154] For the input vector 𝑥, during forward propagation, the original intermediate states are computed. Updated to:

[0155]

[0156] After training, the original parameter matrix and the weights obtained during training and The updated parameter matrix W is calculated as follows:

[0157] ;

[0158] For standard LoRA, the same learning rate for A and B has been shown to lead to suboptimal learning when the embedding dimension is large. In LoRA+, the learning rates for A and B are set as follows:

[0159]

[0160] Let A be the learning rate; Let B be the learning rate; A constant much greater than 1; this mechanism can alleviate the convergence problem under high-dimensional embedding and improve fine-tuning efficiency;

[0161] Loss function of the emotion perception rating agent for:

[0162]

[0163] in For true labels, if the i-th class is a true category, then Otherwise, it is 0; Predict the probability of class i for the model; The logarithm of the predicted probability is taken because when the predicted probability is very close to 1, the loss approaches 0; while when the predicted probability is far from 1, the loss should increase. The emotion perception rating agent is trained for 3 rounds using the training set to obtain the optimal emotion perception rating agent.

[0164] During model training, we determined the following optimal hyperparameters by adjusting the learning rate, the rank value of the low-rank decomposition, and the breakpoints at different stages: learning rate of 0.0001, warm-up rate of 0.1, 3 training epochs, batch size dynamically adjusted according to the model parameter scale, sequence truncation length set to 1024 tokens, AdamW optimizedr selected, and model breakpoints saved every 500 steps.

[0165] This invention aims to improve the scientific rigor and robustness of assessments of negative emotions and stress resilience, reduce subjects' psychological defenses, and enhance assessment efficiency and acceptance in psychological counseling and training settings. The system comprises four modules: model training, topic control, dialogue return, and result output.

[0166] The system is developed using Python, with training data in JSON format. The model training method employs LoRA+ for efficient parameter fine-tuning. By freezing the backbone parameters of the pre-trained model, only the low-rank factorization matrix is ​​updated to improve training efficiency and stability. The loss function is the cross-entropy function.

[0167] In the topic control module (see...) Figure 2 In the system, the initial state is "negative emotion index", and four core topics are evaluated in sequence: negative emotions, anhedonia, physical distress, and high-risk behavior.

[0168] After the system completes the first two topics, if the sum of the two scores ( + If ) = 2, then jump directly to the end node of the stage and terminate the negative emotion detection.

[0169] In the topic of "physical distress", the system first asks a general question, "How is your physical condition recently, and does it affect your life?" If the score is 1, it will move on to the topic of "high-risk behavior". If the score is greater than 1, it will continue to ask sub-questions such as "sleep, appetite, mental fatigue" before moving on to the next topic.

[0170] Once all four topics are completed, the system will store the corresponding questions, answers, and scores in the status for later use.

[0171] In the dialogue return module, the system generates the next question based on the current score and the answer content, performs semantic and sentiment optimization, and outputs a naturally connected transition statement. The result is returned to the front end via HTTP message, where the digital human interaction agent generates voice and facial expressions to achieve natural human-computer dialogue.

[0172] In the results output module, the system calculates the average score for each topic to form a comprehensive negative sentiment index. If the total score is greater than 8, it proceeds to the ruminative thinking test (RRS-10); otherwise, it directly proceeds to the stress resistance test (CD-RISC-10). After the stress resistance test is completed, the system summarizes the scores of each stage and generates a psychological state and stress resistance assessment report.

[0173] In summary, this invention realizes a phased, progressive, and interpretable psychological state assessment system through a multi-agent collaborative mechanism and the LangGraph framework. While improving the scientific rigor and robustness of the assessment, it significantly enhances the user experience and the naturalness of the interaction.

[0174] In a specific embodiment, the system's operation flow mainly includes three core stages:

[0175] Training data preprocessing: The raw corpus is screened, labeled and cleaned, and semantic biases are corrected by combining cognitive psychology rules;

[0176] Model fine-tuning and optimization: The vertical domain model is optimized using efficient parameter fine-tuning methods (such as LoRA+) to ensure efficient generalization and emotion recognition capabilities under limited computing power;

[0177] Multi-agent deployment based on LangGraph: Through a state graph-driven mechanism, task connection and state sharing between agents are realized, completing a closed loop from voice input to multi-dimensional psychological state assessment.

[0178] Step 1: Model fine-tuning and optimization;

[0179] Step 1-1: Training data preprocessing;

[0180] In embodiments of the present invention, the training data mainly comes from multimodal data collection in a psychological counseling room setting, including: (1) the client's voice responses during the counseling process; (2) the scores of the scales filled out simultaneously; and (3) the counseling dialogue text after compliance desensitization processing. To ensure the reliability of subsequent model training, the data preprocessing process includes the following steps:

[0181] First, all speech data is transcribed into text using Whisper, and samples with failed transcriptions or missing values ​​are removed. Second, user responses are mapped to hierarchical labels (Likert four-point scale) based on scale scores, and then encapsulated in a structured JSON format along with the corresponding question text and topic category. This format unifies the questions, answers, labels, and contextual information, facilitating their flow and retrieval in multi-agent environments.

[0182] Building upon this foundation, to address potential cognitive biases in the data (such as discrepancies between visitors' self-assessments and actual performance), the system employs a prompting-based large language model based on cognitive behavioral therapy (CBT) thinking diagnosis for automatic correction, supplemented by manual sampling to improve the accuracy and consistency of annotation. This correction mechanism effectively reduces the interference of recall bias and social desirability effects on sample quality.

[0183] Steps 1-2: Model fine-tuning;

[0184] In the embodiments of the present invention, the base model is selected from the open-source Chinese large-scale pre-trained model ChatGLM4-9B. The present invention adopts an efficient parameter fine-tuning method to optimize the adaptability of the model.

[0185] Specifically, LoRA+ branches are introduced alongside the weight matrices of the attention layer and the feedforward layer. Let the original weight matrix be: given a parameter matrix... Its update form can be expressed as:

[0186]

[0187] in, These are the frozen pre-training parameters. This represents a low-rank approximate update. LoRA+ decomposes it into:

[0188] , , , ≪

[0189] in and This is a low-rank decomposition matrix. During forward propagation, the calculation process for the input vector 𝑥 is as follows:

[0190]

[0191] Matrix A is trained using a larger learning rate, while matrix B is trained using a smaller learning rate. After training, the original parameter matrix is ​​further processed. and the weights obtained during training and Merge: For standard LoRA, the same learning rate for A and B has been shown to lead to suboptimal learning when the embedding dimension is large. In LoRA+, the learning rates for LoRA modules A and B are set to meet optimization requirements:

[0192]

[0193] This mechanism can alleviate the convergence problem under high-dimensional embedding and improve fine-tuning efficiency.

[0194] In terms of training, this invention employs a fine-tuning strategy, using question and answer texts as input sequences to synchronously guide the model in learning the "question-answer-scoring" interaction pattern. The training loss function is cross-entropy.

[0195]

[0196] in For true labels, if the i-th class is a true category, then Otherwise, it is 0. Predict the probability of class i for the model. This is achieved by taking the logarithm of the predicted probability. This is because when the predicted probability is very close to 1, the loss approaches 0. However, when the predicted probability is far from 1, the loss should increase.

[0197] During model training, we determined the following optimal hyperparameters by adjusting the learning rate, the rank of the low-rank decomposition, and the breakpoints at different stages: learning rate of 0.0001, warm-up rate of 0.1, 3 training epochs, batch size dynamically adjusted according to the model parameter size, sequence truncation length set to 1024 tokens, AdamW optimizedr selected, and model breakpoints saved every 500 steps.

[0198] Step 2: System Framework;

[0199] This invention introduces LangGraph as the core orchestration framework at the system implementation level to support the collaborative operation of multiple agents and the orderly scheduling of phased evaluation processes. LangGraph provides a graph-structured state machine mechanism, enabling each functional module to execute independently while also achieving information dependencies and state transfer through explicit data paths.

[0200] Step 2-1: Node definition;

[0201] Each functional module in the system is defined as an independent node. Internally, each node runs its own model inference logic, such as text generation, semantic analysis, speech recognition, or result display. Each node exposes a unified input and output interface for unified scheduling and reuse within the LangGraph framework.

[0202] Step 2-2: Edge and State Transmission;

[0203] The LangGraph framework explicitly describes the data flow between nodes through edges.

[0204] For example, user speech is processed by the transcription node to generate text, which is then passed to subsequent nodes via edges for semantic analysis and score calculation; the score result is then passed to the question planning node to generate the next question.

[0205] All session information (including questions, answers, scores, and stage status) is stored in the global session state to support cross-stage calls and traceability.

[0206] Steps 2-3: Construction of the phased subgraph;

[0207] Negative emotion detection, rumination detection, and stress resilience detection are each constructed as an independent subgraph.

[0208] Each subgraph consists of multiple nodes connected together to complete the task for that stage, and the subgraphs are linked by conditional edges. For example, when the negative emotion score exceeds a threshold, the system automatically jumps to the ruminant thinking detection subgraph to achieve an adaptive assessment branch.

[0209] Steps 2-4: Loop and Termination Conditions;

[0210] The system uses loop edges to automatically iterate the chain of questions. When the termination condition is not met, LangGraph automatically returns to the interaction node to initiate the next round of questions.

[0211] Once all four core topics have been scored, the system determines the end of this phase and persistently saves the results. Through the above design, this invention forms a technical path centered on "node scheduling—graph structure control—phase progression." Compared with traditional linear questionnaire methods, this architecture can dynamically adjust the dialogue path, maintain a closed loop of interaction logic, and possess higher scalability and robustness.

Claims

1. A progressive stress resistance assessment system based on a large-scale intelligent agent model, characterized in that, Includes the following steps: Step S1: Construct the dataset; We collected the following raw data: the voice responses of visitors during consultations, the PHQ-9 scale scores completed by visitors, and the consultation dialogue text after compliance-compliant anonymization. We then used Whisper to transcribe the voice responses from the raw data into text, removing samples with transcription failures and missing values. Based on the PHQ-9 scale scores, we mapped user responses to hierarchical labels. We then used the hierarchical labels, question text, response text, and topic category as a set of data. Based on these sets of data, we constructed a training set. Step S2: Construct a multi-agent cooperative network; the multi-agent cooperative network is constructed based on the LangGraph framework; the multi-agent cooperative network is a state machine; The multi-agent collaborative system consists of several nodes and edges between them. Nodes achieve data flow and state transfer through edges. The nodes in the multi-agent collaborative network include a digital human interaction agent, an emotion perception and scoring agent, an emotion guidance and planning agent, an empathy expression generation agent, a scale scoring agent, a state object, and a question list. The emotion perception and scoring agent is trained using a dataset. When a predetermined training condition is met, the optimal emotion perception and scoring agent is obtained. The optimal emotion perception and scoring agent is then used to replace the existing emotion perception and scoring agent in the multi-agent collaborative network. Step S3: Use a multi-agent cooperative network to detect the user's stress resistance and obtain a stress resistance index. If the stress resistance index A stress tolerance index of 40 or higher indicates a user with strong resilience; if the stress tolerance index is higher... If the stress tolerance index is greater than or equal to 30 and less than 40, then the user's stress tolerance is within the normal range; if the stress tolerance index... If the value is less than 30, the user's stress tolerance is relatively weak.

2. The progressive stress resistance assessment system based on a large-scale intelligent agent according to claim 1, characterized in that, The digital human interactive agent is constructed based on the ER-NeRF model; the input of the digital human interactive agent is speech text; the digital human interactive agent outputs natural speech and facial expressions through TTS speech synthesis method, presenting questions, empathic feedback and scale items to the user; The emotion perception and scoring agent is based on the vertical domain large model ChatGLM4-9B, which is efficiently fine-tuned with LoRA+ parameters. The input of the emotion perception and scoring agent is the question text and the user's answer text. The output of the emotion perception and scoring agent is the negative emotion type and the negative emotion score. The negative emotions include the negative emotion index, the anhedonia index, the physical distress index, and the high-risk behavior index. The negative emotion score is in the range of [1, 4]. The emotion perception and scoring agent writes the negative emotion score into the state object according to the negative emotion type using the state update method. The negative emotion index S_NE (Negative Emotion Sub-Index) reflects the intensity of expressions related to low mood; the anhedonia index S_ANH (Anhedonia Sub-Index) reflects the degree of reduced interest / pleasure in daily activities; the somatic distress index S_SOM (Somatic Distress Sub-Index) reflects the intensity of physiological discomfort related to mental and physical stress; and the risk behavior index S_RISK (Risk Behavior Sub-Index) reflects the intensity of potential high-risk behaviors such as impulsivity / risk-taking. The emotion-guided planning agent is based on DeepSeek-R1; The inputs to the emotion-guided planning agent are the current question, the user's response text, and negative emotion indices, anhedonia indices, physical distress indices, and high-risk behavior indices. The output of the emotion-guided planning agent is a question text; The emotion-guiding planning agent checks whether all questions in the current index category in the question list have been asked. If all questions have been asked, it proceeds to the next index category. If the question is not completed, the next question is drawn from the index category of the question list; Iterate through all index categories of questions until all questions in the question list have been asked. Specifically, in the Physical Distress Index category, if the first question is "How is your physical condition recently, and does it affect your life?", if the score for the first question is 1, it means that the user has not experienced any obvious physical discomfort, and the user is moved to the next index category; if the score for the first question is > 1, then the other questions in the index category are traversed. If the current index is negative sentiment, the next index category will be anhedonia index. If the current index is the anhedonia index, calculate the sum of the user's scores on the negative emotion index and the anhedonia index. If the sum of the scores on the negative emotion index and the anhedonia index is less than or equal to 2, it is determined that the current stage does not meet the screening criteria for depressive episodes; if the sum of the scores on the negative emotion index and the anhedonia index is greater than 2, the next index category is the physical distress index. The empathy expression generation agent is based on DeepSeek-R1; the input of the empathy expression generation agent is the current question text, the user's answer text, and the next question text; the output of the empathy expression generation agent is the natural response text, which is then used as speech text and input into the digital human interaction agent. The scale scoring and management agent includes a ruminant index. Calculation module and stress resistance index Calculation module; Rumination index The calculation module calculates the Rumination Index based on the scores of questions in the Rumination RRS-10 scale. Compression resistance index The calculation module calculates the resilience index based on the scores of questions in the simplified psychological resilience scale CD-RISC-10. ; Rumination Index With the state object Perform summation, and use the result of the summation as the new... Write to the state object; set the resilience index With the state object Perform summation, and use the result of the summation as the new... Write the state object; the input for the scale rating and management agent is the negative emotion type and negative emotion score; The state object is: };in, The index categories represent the current issues; these categories include negative emotion index, anhedonia index, physical distress index, high-risk behavior index, rumination index, and stress resilience index. This represents the negative sentiment index value. This represents the state value of the anhedonia index. The physical distress index is the state value. This represents the risk index status value for high-risk behaviors. The total score represents negative emotions. This represents the rumination index state value. This represents the compressive strength index state value. The state update method is as follows: If the negative emotion type is a negative emotion index, then accumulate the values ​​in the state object. The value will be added to the state object. The value is divided by the total number of questions under the negative sentiment index question category; If the negative emotion type is the anhedonia index, then the index is accumulated in the state object. The value will be added to the state object. The value is divided by the total number of problems under the Anhedonia Index problem category; If the negative emotion type is the physical distress index, then accumulate the values ​​in the state object. The value will be added to the state object. The value is divided by the total number of problems under the Physical Distress Index problem category; If the negative emotion type is a high-risk behavior risk index, then the index is accumulated in the state object. The value will be added to the state object. The value is divided by the total number of questions under the high-risk behavior risk index question category; The list of questions is based on the MINI scale; The issues in the three items of depressive episode, dysphoric mood and suicide in the MINI scale were reclassified into four dimensions according to the negative emotion index, anhedonia index, physical distress index and high-risk behavior risk index. The negative emotion index dimension includes questions such as "How have you been feeling lately, and has it affected your life?" and "Have you felt depressed or down for most of the day in the past two weeks?" Questions for the anhedonia index dimension include "Do you usually enjoy doing things, and have you done any of these things recently?" and "What types of TV programs do you usually like to watch?" Questions in the Physical Distress Index dimension include: "How is your physical condition lately, and has it affected your life?", "What do you want to do when you can't fall asleep?", "Have you had a poor appetite or overeaten lately?", "Do you feel tired and lack energy every day?", "Have you recently had difficulty concentrating, hesitated, or had difficulty making decisions?" The high-risk behavior risk index includes questions such as "What are your views on suicide?" and "Have you recently had recurring thoughts of self-harm, suicide, or wishing to die?" By default, multi-agent cooperative networks start with questions about "negative emotions".

3. The progressive stress resistance assessment system based on a large-scale intelligent agent according to claim 1, characterized in that, The steps for detecting a user's resilience using a multi-agent cooperative network are as follows: Step S3-1: Detect the negative emotion stage; Step S3-1-1: The digital human interactive intelligent agent outputs the first question; The emotion-guided planning agent selects question text from the question list; If this is the first time asking a question, start with a question related to "negative emotions" and get a voice response from the user. If this is not the first time a question has been asked, retrieve it from the state object. Proceed to step S3-1-4 to obtain the next question text; Step S3-1-2: Based on the user's voice response, use the Whisper speech recognition model to transcribe the user's voice response into speech text; Step S3-1-3: Analysis and quantification of the agent for emotion perception scoring; Input the voice text and the corresponding question text into the emotion perception and scoring agent; the emotion perception and scoring agent outputs a negative emotion score based on the type of negative emotion. Step S3-1-4: The emotion-guided planning agent generates the next question; Input the negative emotion score into the emotion-guided planning agent; the emotion-guided planning agent outputs the text of the next question; Step S3-1-5: Express empathy, generate an intelligent agent to optimize the response, and provide feedback to the user; Input the text of the next question in S3-1-4 into the empathic expression to generate an agent and obtain a natural response text; The logic for generating natural response text includes an emotional feedback layer and a guidance layer. The emotional feedback layer responds to the user's answer, demonstrating understanding and emotional support. The guidance layer embeds a transitional phrase for the next question at the end of the response to maintain contextual coherence and emotional rhythm. The natural response text is input into the digital human interaction agent for speech synthesis and facial expression driving, achieving a unified semantic and visual empathetic communication experience through streaming video. If the natural response text does not contain the next question, proceed to step 2; if the natural response text does contain the next question, proceed to step 1-1. Step S3-2: Rumination detection phase; The total negative affect score is calculated as the Negative Affect Total Score. :

4. Among them The index score for the k-th topic; When the total score for negative emotions If the score is less than the preset threshold of 8 points, proceed to step 3; When the total score for negative emotions When the score exceeds a preset threshold of 8 points, the digital human interaction agent presents the user with RRS-10 questions. The user answers by clicking on the screen, avoiding the additional cognitive load of voice input; Rumination Index The calculation module calculates the ruminant thinking index based on the user's clicks and answers. ; Step S3-3: Compression resistance testing stage; The digital human interactive agent presents users with questions from the CD-RISC-10 scale, and users respond using a screen-clicking mode; stress resilience index. The calculation module calculates the stress resistance index based on the user's clicks and answers. ; Step S3-4: Summarize the results; If the stress resistance index A stress tolerance index of 40 or higher indicates a user with strong resilience; if the stress tolerance index is higher... If the stress tolerance index is greater than or equal to 30 and less than 40, then the user's stress tolerance is within the normal range; if the stress tolerance index... If the value is less than 30, the user's stress tolerance is relatively weak.

5. The progressive stress resistance assessment system based on a large-scale intelligent agent according to claim 2, characterized in that, The fine-tuning process for the emotion perception and scoring agent is as follows: The parameter matrix of the large vertical domain model ChatGLM4-9B is decomposed using LoRA+: The parameter matrix is , The update format is as follows:

6. Among them, These are the frozen pre-training parameters. This indicates a low-rank approximate update; Use LoRA+ to The solution is: ， ， ， ≪ ； in and All are low-rank decomposition matrices; for The rank of the matrix after low-rank decomposition; for The dimension; Original calculation intermediate state for:

7. For the input vector 𝑥, during forward propagation, the original intermediate computation states are... Updated to:

8. After training is complete, convert the original parameter matrix... and the weights obtained during training and The updated parameter matrix W is calculated as follows: ； For standard LoRA, the same learning rate for A and B has been shown to lead to suboptimal learning when the embedding dimension is large. In LoRA+, the learning rates for A and B are set as follows:

9. Let A be the learning rate; Let B be the learning rate; A constant much greater than 1; this mechanism can alleviate the convergence problem under high-dimensional embedding and improve fine-tuning efficiency.

10. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the progressive stress resistance assessment system based on a large model intelligent agent as described in any one of claims 1-4.

11. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the progressive stress resistance assessment system based on a large model intelligent agent as described in any one of claims 1-4.

Images