AI psychological counseling method and system

Abstract

The invention discloses an AI psychological counseling method and system, and the method comprises the steps: collecting the physiological signal data, behavior signal data, text signal data and individual attribute data of an individual, and the individual attribute data at least comprises the gender; extracting physiological features from the physiological signal data, extracting behavior features from the behavior signal data, extracting text features from the text signal data, and determining the individual attribute data based on the physiological features, the behavior features and the text features in combination with the individual attribute data. Respectively calculating a depression risk value, an anxiety risk value, a pressure load risk value and a crisis warning risk value to form a four-dimensional theme risk vector; calculating a total risk value according to the four-dimensional theme risk vector; and comparing the total risk value with a preset threshold sequence, and determining individual risk levels including a health level, a mild level, a moderate level, a severe level and a crisis level. According to the method and the system, the technical problem of large single-mode error in the traditional psychological service is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and mental health service technology, and in particular to an AI-based psychological counseling method and system. Background Technology

[0002] Currently, technological development in the field of mental health services mainly focuses on the following aspects: First, at the level of psychological state assessment, existing technologies largely rely on unimodal or bimodal data analysis. The most common approach is online questionnaires based on self-report scales, which are highly subjective and prone to masking. In addition, some technologies employing voice emotion recognition or facial expression analysis have emerged, attempting to objectively assess emotional states through a single behavioral channel. However, psychological state is a complex manifestation of the combined effects of physiological, behavioral, cognitive, and environmental factors. The aforementioned unimodal or bimodal methods have inherent limitations: First, the assessment accuracy is insufficient; relying solely on questionnaires or single behavioral characteristics easily leads to false positives (misjudging normal emotional fluctuations as illness) or false negatives (missing the true risk), making it difficult to meet the accuracy requirements of large-scale screening. Second, the decision-making process is uninterpretable; many deep learning-based models are like "black boxes," only outputting results without providing convincing decision-making evidence, making their results difficult for counselors to accept or hindering regulatory audits.

[0003] Secondly, at the intervention and treatment level, existing technologies generally suffer from a "broken screening-intervention link." Most systems only complete a "one-time assessment" and generate a static report, while subsequent steps such as consultation appointments, resource matching, interview recording, and quality supervision still heavily rely on manual operation. This break leads to a series of operational bottlenecks: assessment results "sleep," failing to automatically trigger effective intervention actions; manual scheduling is time-consuming and inefficient, unable to achieve minute-level rapid response; the consultation process lacks real-time, quantitative quality control methods, and supervision is severely lacking. The entire service process fails to form an automated closed loop of "assessment-decision-treatment-quality control," severely restricting service efficiency and coverage.

[0004] Furthermore, existing methods have extremely limited capabilities in long-term support and group intervention. Current AI chatbots mostly use fixed scripts, making it difficult to conduct dynamic and personalized deep interactions based on individual risk profiles. This often results in irrelevant answers, poor user experience, and low long-term compliance. At the same time, current technologies lack the ability to identify common weaknesses from group data, hindering contextualized group training tailored to individual needs. This makes interventions less targeted and significantly reduces their effectiveness.

[0005] In summary, existing technologies face three major technical challenges: insufficient accuracy and interpretability in single-modal assessments, a broken link in the screening-intervention-supervision process, and a lack of long-term support and group intervention capabilities. Therefore, there is an urgent need in this field for an integrated solution that can fuse multimodal data to achieve high-precision, interpretable risk quantification, and automatically connect the entire process of screening, scheduling, interviews, supervision, and long-term support. Summary of the Invention

[0006] To address the technical problems existing in the background art, this invention proposes an AI-based psychological counseling method and system.

[0007] The present invention proposes an AI-based psychological counseling method, comprising the following steps: S1. Collect physiological signal data, behavioral signal data, text signal data, and individual attribute data of an individual, wherein the individual attribute data includes at least gender; S2. Extract physiological features from physiological signal data, behavioral features from behavioral signal data, and text features from text signal data. Based on physiological features, behavioral features, and text features, and in conjunction with the individual attribute data, calculate depression risk value, anxiety risk value, stress load risk value, and crisis warning risk value respectively to form a four-dimensional subject risk vector. S3. Calculate the total risk value based on the four-dimensional thematic risk vector; S4. Compare the total risk value with a preset threshold sequence to determine the individual risk level, which includes health level, mild level, moderate level, severe level and crisis level; S5. Automatically match consulting resources based on individual risk levels and generate scheduling plans and preliminary screening reports.

[0008] Preferably, the physiological signal data includes at least photoplethysmography (PPG) signals and blood oxygen saturation signals acquired by a finger-clip PPG sensor, and electroencephalogram (EEG) signals acquired by a wearable EEG sensor; the behavioral signal data includes at least speech signals acquired by a microphone and facial video signals acquired by a camera; the text signal data is obtained by converting the speech signals into text through speech recognition, or by directly receiving text data input by the user.

[0009] Preferably, the extraction of physiological features from physiological signal data specifically includes: The root mean square values ​​of heart rate and the difference between adjacent heartbeats are extracted from photoplethysmography (PPG) signals; the standard deviation of blood oxygen saturation is extracted from blood oxygen saturation signals; and the standard deviation of blood oxygen saturation is extracted from electroencephalogram (EEG) signals. Bandwidth power and The ratio of bandwidth power; The extraction of behavioral features from behavioral signal data specifically includes: Speech rate features and fundamental frequency features are extracted from speech signals, and the average intensity of frowning occurrences of facial frowning action units is extracted from facial video signals. The extraction of text features from text signal data specifically includes: By performing semantic analysis on the text using a large language model, the text can be directly output as a risk value for depression, anxiety, stress, and crisis.

[0010] Preferably, the depression risk value is obtained by: selecting a corresponding weighted calculation model based on gender information, and calculating the text depression risk value, the root mean square value of the difference between adjacent heartbeat intervals, the fundamental frequency feature, the standard deviation of blood oxygen saturation, and... Bandwidth power and The ratio of bandwidth power is weighted and the calculation result is limited to a predetermined range. The anxiety risk value is obtained by weighted summation of text anxiety risk value, speech rate characteristics, and average intensity of frowning, and the summation result is limited to a predetermined numerical range.

[0011] Preferably, the stress load risk value is obtained by weighted summing of the text stress risk value and the physiological stress index calculated based on the root mean square value of the difference between adjacent heartbeat intervals, and limiting the summation result to a predetermined numerical range. The value of the crisis warning risk value is directly determined by the risk value of the text crisis level, and when a preset high-risk intent keyword is identified, the crisis warning risk value is set to the defined maximum value.

[0012] Preferably, the semantic analysis of the text using a large language model specifically includes: The text obtained from speech recognition or the text data directly input by the user, along with the preset psychological analysis prompts, are input into the large language model. The large language model is used to perform semantic analysis on the text and directly output four independent text risk values ​​in parallel. The text risk values ​​include: text depression risk value, text anxiety risk value, text stress risk value, and text crisis level risk value. The large language model is configured to simultaneously evaluate and output a quantitative value of the risk level of the text in four psychological dimensions: depression, anxiety, stress, and crisis, based on the input text content. Each quantitative value is a value between 0 and 1.

[0013] Preferably, step S3 specifically includes: The four-dimensional subject risk vector, composed of the depression risk value, anxiety risk value, stress load risk value and crisis warning risk value, is multiplied by a preset weight vector, and the result of the dot product operation is the total risk value. The preset weight vector contains configurable weight coefficients corresponding to each risk value in the four-dimensional subject risk vector.

[0014] Preferably, step S4 specifically includes: The total risk value is compared to a preset threshold sequence consisting of five increasing thresholds; Among them, when the total risk value falls between the first threshold and the second threshold, it is judged as a healthy level; When the total risk value falls between the second and third thresholds, it is classified as a mild level. When the total risk value falls between the third and fourth thresholds, it is classified as a moderate level. When the total risk value falls between the fourth and fifth thresholds, it is classified as severe. When the total risk value is greater than or equal to the fifth threshold, it is judged to be at the crisis level.

[0015] Preferably, step S5 specifically includes: Based on the preset treatment mapping rules, the corresponding type of consultation resources is matched to the individual risk level; The handling mapping rules include at least the following: mapping health level to targeted training resources, mapping mild and moderate levels to AI interview resources, mapping severe levels to human consultation resources with additional AI supervision resources, and mapping crisis level to human consultation resources and initiating crisis handling procedures. Based on the matched consultation resource type, query the corresponding resource pool status, and generate a scheduling plan containing specific time, consultant ID and contact channel based on the priority scheduling algorithm; Meanwhile, based on individual risk levels, total risk values, and four-dimensional thematic risk vectors, a preliminary screening report is generated through a large model, which includes a summary of objective indicators, an interpretation of risk factors, and suggested handling paths. For severe and crisis levels, the initial screening report will simultaneously generate a work order for manual review.

[0016] Preferably, it also includes: S6, conducting highly empathetic structured interviews based on individual risk levels and initial screening reports, generating interview transcripts, interview highlights, and structured interview reports.

[0017] This invention proposes an AI-based psychological counseling system, comprising: The data acquisition module is used to collect physiological signal data, behavioral signal data, text signal data, and individual attribute data of an individual, wherein the individual attribute data includes at least gender; The first processing module is used to extract physiological features from physiological signal data, behavioral features from behavioral signal data, and text features from text signal data. Based on the physiological features, behavioral features, and text features, and in combination with the individual attribute data, it calculates depression risk value, anxiety risk value, stress load risk value, and crisis warning risk value respectively to form a four-dimensional subject risk vector. The second processing module is used to calculate the total risk value based on the four-dimensional subject risk vector; The third processing module is used to compare the total risk value with a preset threshold sequence to determine the individual risk level, which includes a health level, a mild level, a moderate level, a severe level, and a crisis level. The data output module is used to automatically match consultation resources based on individual risk levels and generate scheduling plans and preliminary screening reports.

[0018] This invention proposes an AI-based psychological counseling method and system. By integrating physiological, behavioral, and textual features through a multimodal screening agent, it constructs an interpretable four-dimensional psychological risk vector, achieving a technological leap from single subjective judgment to multi-source objective quantification of psychological states, significantly improving screening accuracy and reliability. A one-click referral scheduling mechanism based on risk grading establishes an automated link from assessment to intervention, enabling intelligent scheduling of counseling resources and minute-level response, completely resolving the pain point of disconnect between screening and treatment in traditional processes. The personalized AI interview and supervision module, through highly empathetic dialogue and real-time ethical verification, forms a personalized service with a "one person, one plan" approach and a closed-loop quality control throughout the entire process, greatly improving the professionalism and efficiency of intervention. The continuous companionship and targeted training engine, through dynamic difficulty adjustment and contextualized training, achieves a dual-cycle intervention from individual to group, effectively ensuring long-term adherence and overall resilience of psychological services. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the workflow of an AI-based psychological counseling method proposed in this invention. Figure 2 This is a schematic diagram illustrating the implementation process of an AI-based psychological counseling method proposed in this invention. Figure 3 This is a schematic diagram of the system architecture of an AI psychological counseling system proposed in this invention. Detailed Implementation

[0020] Reference Figures 1-2 The present invention proposes an AI-based psychological counseling method, comprising the following steps: S1. Collect physiological signal data, behavioral signal data, text signal data, and individual attribute data of individuals, including at least gender.

[0021] In this embodiment, the physiological signal data includes at least the photoplethysmography (PPG) signal and blood oxygen saturation signal acquired by a finger-clip PPG sensor, and the electroencephalogram (EEG) signal acquired by a wearable EEG sensor; the behavioral signal data includes at least the speech signal acquired by a microphone and the facial video signal acquired by a camera; and the text signal data is acquired by converting the speech signal into text through speech recognition, or by directly receiving text data input by the user.

[0022] S2. Extract physiological features from physiological signal data, behavioral features from behavioral signal data, and textual features from textual signal data. Based on the physiological features, behavioral features, and textual features, and in conjunction with the individual attribute data, calculate the depression risk value, anxiety risk value, stress load risk value, and crisis warning risk value respectively to form a four-dimensional subject risk vector.

[0023] In this embodiment, physiological features are extracted from physiological signal data, specifically including: The root mean square values ​​of heart rate and the difference between adjacent heartbeats are extracted from photoplethysmography (PPG) signals; the standard deviation of blood oxygen saturation is extracted from blood oxygen saturation signals; and the standard deviation of blood oxygen saturation is extracted from electroencephalogram (EEG) signals. Bandwidth power and The ratio of bandwidth power; Extracting behavioral features from behavioral signal data specifically includes: Speech rate features and fundamental frequency features are extracted from speech signals, and the average intensity of frowning occurrences of facial frowning action units is extracted from facial video signals. Extracting text features from text signal data specifically includes: By performing semantic analysis on the text using a large language model, the text can be directly output as a risk value for depression, anxiety, stress, and crisis.

[0024] In this embodiment, the depression risk value is obtained as follows: based on gender information, a corresponding weighted calculation model is selected, and the text depression risk value, the root mean square value of the difference between adjacent heartbeat intervals, the fundamental frequency feature, the standard deviation of blood oxygen saturation, and the... Bandwidth power and The ratio of bandwidth power is weighted and the calculation result is limited to a predetermined range.

[0025] It should be noted that the predetermined numerical range is a closed interval from 0 to 1.

[0026] Specifically, when the user is male, the formula for calculating the depression risk score is as follows: When the user is female, the formula for calculating the depression risk score is as follows:

[0027] in, The text represents the risk value for depression. This is the root mean square value of the difference between adjacent heartbeat intervals; It is a fundamental frequency characteristic; The standard deviation of blood oxygen saturation; This represents the risk value for depression. for Bandwidth power and The ratio of bandwidth power; Weights associated with the text's depression risk score; The weights associated with the root mean square value of the difference between adjacent heartbeat intervals; Weights related to the fundamental frequency characteristics; Weights related to the standard deviation of blood oxygen saturation; To and Bandwidth power and The weighting related to the ratio of bandwidth power; and .

[0028] In this embodiment, the anxiety risk value is obtained by weighted summation of the text anxiety risk value, speech rate features, and average intensity of frowning, and the summation result is limited to a predetermined numerical range.

[0029] Specifically, the formula for calculating the anxiety risk value is as follows: ; in, Text anxiety risk value; Characteristic of speech rate; The average intensity of a frown; Anxiety risk value; The weights associated with the text anxiety risk value; The correlation weights with speech rate features; The weights associated with the average intensity of frowning, and .

[0030] In this embodiment, the stress load risk value is obtained by weighted summing of the text stress risk value and the physiological stress index calculated based on the root mean square value of the difference between adjacent heartbeat intervals, and the summation result is limited to a predetermined numerical range.

[0031] Specifically, the formula for calculating the pressure load risk value is as follows: ; in, Text stress risk value; This represents the pressure load risk value. Weights associated with text stress risk values; The weights are related to the root mean square value of the pressure load risk value and the difference between adjacent heartbeat intervals, and ; It is the root mean square value of the difference between adjacent heartbeat intervals.

[0032] In this embodiment, the value of the crisis warning risk value directly depends on the risk value of the text crisis level, and when a preset high-risk intent keyword is identified, the crisis warning risk value is set to the defined maximum value.

[0033] Specifically, the crisis warning risk value depends directly on the text features that represent the crisis tendency, i.e., the text crisis level risk value, and when a preset high-risk intent keyword is identified, the crisis warning risk value is set to the maximum value.

[0034] In this embodiment, semantic analysis of the text is performed using a large language model, specifically including: The text obtained from speech recognition or the text data directly input by the user, along with the preset psychological analysis prompts, are input into the large language model. By performing semantic analysis on the text using a large language model, four independent text risk values ​​are directly and in parallel output. These text risk values ​​include: text depression risk value, text anxiety risk value, text stress risk value, and text crisis level risk value. The large language model is configured to simultaneously evaluate and output the risk level of the text in four psychological dimensions—depression, anxiety, stress, and crisis—based on the input text content, with each value ranging from 0 to 1.

[0035] S3. Calculate the total risk value based on the four-dimensional thematic risk vector.

[0036] In this embodiment, step S3 specifically includes: The four-dimensional thematic risk vector, composed of depression risk value, anxiety risk value, stress load risk value and crisis warning risk value, is multiplied by a preset weight vector. The result of the dot product operation is the total risk value. The preset weight vector contains configurable weight coefficients corresponding to each risk value in the four-dimensional subject risk vector.

[0037] Specifically, the formula for calculating the total risk value is as follows: ; in, This represents the risk value for depression. Anxiety risk value; This represents the pressure load risk value. Risk value for crisis warning; Total risk value; Weights associated with depression risk scores; Weights associated with anxiety risk scores; Weights associated with pressure load risk values; The weights associated with the risk value of the crisis warning, and .

[0038] S4. Compare the total risk value with the preset threshold sequence to determine the individual risk level. The individual risk level includes health level, mild level, moderate level, severe level and crisis level.

[0039] In this embodiment, step S4 specifically includes: The total risk value is compared to a preset threshold sequence consisting of five increasing thresholds; Among them, when the total risk value falls between the first threshold and the second threshold, it is judged as a healthy level; When the total risk value falls between the second and third thresholds, it is classified as a mild level. When the total risk value falls between the third and fourth thresholds, it is classified as a moderate level. When the total risk value falls between the fourth and fifth thresholds, it is classified as severe. When the total risk value is greater than or equal to the fifth threshold, it is judged to be at the crisis level.

[0040] S5. Automatically match consulting resources based on individual risk levels and generate scheduling plans and preliminary screening reports.

[0041] In this embodiment, step S5 specifically includes: Based on the preset treatment mapping rules, the corresponding type of consultation resources is matched to the individual risk level; Among them, the handling mapping rules include at least: mapping health level to targeted push training resources, mapping mild and moderate levels to AI interview resources, mapping severe level to human consultation resources and attaching AI supervision resources, and mapping crisis level to human consultation resources and initiating crisis handling process. Based on the matched consultation resource type, query the corresponding resource pool status, and generate a scheduling plan containing specific time, consultant ID and contact channel based on the priority scheduling algorithm; Meanwhile, based on individual risk levels, total risk values, and four-dimensional thematic risk vectors, a preliminary screening report is generated through a large model, which includes a summary of objective indicators, an interpretation of risk factors, and suggested handling paths. For severe and crisis levels, the initial screening report will simultaneously generate a work order for manual review.

[0042] Specifically, when an individual's risk level is health level L0, targeted training resources are pushed to them; when an individual's risk level is mild level L1 or moderate level L2, they are mapped to AI interview resources; when an individual's risk level is severe level L3, they are mapped to human priority + AI supervision; when an individual's risk level is crisis level L4, they are mapped to human consultation resources and the crisis escalation and hospital referral procedures are initiated.

[0043] In this embodiment, objective indicators are interpreted using a large model and specially designed prompts to form "objective indicators + explainable causes + suggested paths." Suggestions for coaching, companionship, and interviews are generated for L1-L2 levels; and action orders are generated for L3-L4 levels for manual review and management. Simultaneously, intelligent scheduling and referral are performed based on the grading results and resource pool (available consultants, time slots, and channels), with full traceability. During scheduling, the target, action path, and expected outcomes are generated into scheduling entries and a traceable queue, supporting priority rules (crisis > severe > moderate > mild). One-click appointments or referrals are available for AI coaching / companionship / interviews / human consultations / hotlines; to prevent harassment, cooldown time windows and concurrency limits are set.

[0044] In this embodiment, it also includes: S6, conducting highly empathetic structured interviews based on individual risk levels and initial screening reports, and generating interview transcripts, interview highlights, and structured interview reports.

[0045] Specifically, a standardized psychological interview is conducted using a highly empathetic style, collecting information based on 30 frequently asked questions by counselors and AI screening results; a structured AI interview report is automatically generated.

[0046] It should be noted that the highly empathetic structured interview utilizes the large model Prompt (interview execution), specifically set as follows: Role: You are a psychotherapist with high empathy and structured interviewing skills.

[0047] Objective: To cover 30 frequently asked questions in a gentle, neutral, and non-judgmental tone, sequentially or as needed, with 1-2 follow-up questions if necessary; to provide a "skip" option for sensitive questions. Upon encountering high-risk semantics, stop follow-up questions and trigger a safety loop. After each question, wait for the user's response and record the verbatim transcript and key points.

[0048] Output: Three objects are continuously maintained in the background of the session: transcript, key points, and report.

[0049] Termination criteria: Coverage of 30 questions ≥ 80% or the user wishes to terminate; a structured interview report and recommended actions will then be generated.

[0050] In this embodiment, it also includes: using an AI-supervised intelligent agent to perform real-time ethical compliance verification on the real-person consultation process based on a structured interview report, generating sentence-by-sentence suggestions and an overall score; using a companion and question-solving intelligent agent to provide daily companionship services based on the individual's risk level, and to execute dialogue strategies such as empathic reflection, interpretation clarification, attribution reconstruction, micro-goal setting, and debriefing reminders; and using a targeted push training intelligent agent to generate virtual scenario role-playing training based on statistical results of the group's four-dimensional topic risk vector, and to dynamically adjust the training difficulty.

[0051] Specifically, for real-person psychological counseling, the purpose of AI supervision is to assess the counselor's capabilities and provide professional advice. Therefore, the client's AI screening results serve as pre-training material for the AI ​​supervisor, enabling full-process observation throughout the counseling process. Based on the general framework and ethical guidelines for psychological counseling supervision (boundaries, informed consent, confidentiality, risk prioritization, non-judgment, and multicultural sensitivity), AI supervision suggestions are generated for experts to review and verify with a single click. Each AI / councilor statement generates a suggestion and compliance verification tag, and an overall score is compiled, generating a verbatim transcript, sentence-by-sentence suggestions, an overall score, and improvement points. A real-time safety sentinel function is also provided: taboo word and boundary-crossing pattern recognition, high-risk intent detection, and alerts.

[0052] In this embodiment, targeting a group (school / community / enterprise), based on the entire chain of data from screening, companionship, interviews, and supervision, the system identifies commonly weak psychological capacity dimensions. Weekly practice sessions with mini-lessons and virtual scenario exercises are generated to improve the group's psychological resilience and coping skills. The weakness dimension mining first calculates the mean and hit rate of each dimension based on the four-dimensional risk vector statistics of the group and individuals, and then uses keyword co-occurrence networks from interviews and companionship dialogues, along with supervisory markers (such as "avoidance" and "over-labeling"), to pinpoint high-frequency weaknesses. Subsequently, clustering methods are used to group members within the target group, resulting in actionable population profiles. A rule supplement layer is then introduced to filter candidate topics according to the policy whitelists and blacklists of schools / communities / enterprises, ensuring compliance and suitability. In the topic generation stage, a topic library (such as "exam anxiety," "public speaking," and "interdepartmental conflict") is used for enhanced matching, and then large-scale model prompts are used to generate AI character dialogue to create virtual scenarios. The virtual scenario exercise unfolds through conversational role-playing: the user plays "I," while the intelligent agent plays "leader / classmate / family member." The system dynamically adjusts follow-up questions and emotional intensity based on feedback, hesitation duration, and error rate. The entire process incorporates built-in content security and ethical constraints: topic intensity thresholds, blacklisted words, and a mechanism for "exiting / skipping questions" at any time to prevent secondary harm.

[0053] In this embodiment, the role settings of the large model system related to the AI ​​supervisory agent are constrained as follows: You are a senior psychological counseling supervisor, familiar with supervision techniques, ethical guidelines (informed consent, confidentiality, boundaries, risk prioritization, multicultural sensitivity, non-judgmental attitude), and key points of structured interviews. Your task is to supervise and assess the quality of the entire conversation, generating sentence-by-sentence suggestions and a post-meeting summary. It is imperative to be objective, neutral, and respectful of the client; prioritize identifying risk scenarios; and output results that strictly conform to the specified format.

[0054] In this embodiment, highly empathetic and accessible daily emotional support, psychological Q&A, and knowledge explanations are provided to individual clients. The system automatically retrieves their screening (L0-L4), interview reports, and supervisory suggestions as context, ensuring continuous, personalized support. This highly empathetic support dialogue aims to provide continuous and gentle emotional support and psychological assistance to individual clients. On one hand, it includes a built-in scale popularization and knowledge base, as well as plain-language explanations of common psychological terms, along with immediately applicable coping techniques (such as breathing exercises, behavioral activation, and exposure-response prevention). On the other hand, it combines dynamic information from the individual (including the client's risk level, risk vector, key event timeline, and the desensitized, publicly available portion of supervisory suggestions) with a dialogue strategy of "empathic reflection → interpretation and clarification → attribution reconstruction → micro-goal setting → review and reminders" to generate a personalized and actionable help path in real time. The specific working method of this embodiment is as follows: You are a health assistant who provides continuous support. Prioritize expressing empathy and use concise, non-judgmental language to communicate with the visitor. You are already aware of the visitor's profile and summaries of their recent screenings, interviews, and supervision. This information is only used for personalized expression and matching suggestions, and it is strictly forbidden to expose any specific identifiable private details. When the visitor asks about scales or terms, please provide easy-to-understand explanations and offer 1-3 actionable steps. When high-risk expressions are detected (identified by a dictionary and classifier), you must immediately trigger the safety handling process, first stabilize the emotions, and then provide localized emergency resources and help channels. Each response should clearly state the next steps and support setting reminders for subsequent review and continuous follow-up.

[0055] Example 1: Physiological signal data of user A (male) is collected through wearable hardware. Specifically, the root mean square (RMSSD) of the difference between adjacent heartbeats and the heart rate (HR) are calculated using the waveform output by a finger clip-on photoplethysmography (FPPC) sensor; the standard deviation of blood oxygen saturation is calculated by outputting the corresponding blood oxygen saturation value per second from a blood oxygen saturation sensor. Brainwaves are collected using an electroencephalogram (EEG) sensor. "Energy and" "Energy, based on brainwaves" "Energy and" "Energy calculation" Bandwidth power and The ratio of bandwidth power.

[0056] User-related behavioral and textual signal data are extracted by collecting voice data, facial video data, and user input data. Speech rate features and fundamental frequency features are extracted from the voice data, where the fundamental frequency feature is the highest frequency in the collected voice data. The average intensity of frowning actions is extracted from the facial video signals; specifically, the average intensity is the proportion of frowning frames to total video frames within the acquisition time.

[0057] By performing semantic analysis on the text using a large language model, the text can be directly output as a risk value for depression, anxiety, stress, and crisis.

[0058] Specifically, the physiological signal data collected in this embodiment includes: the root mean square (RMSSD) of the difference between adjacent heartbeat intervals (28) and the standard deviation of blood oxygen saturation. 0.25 and Bandwidth power and The ratio of frequency band power is 0.58; semantic analysis of the text using a large language model yields the following results: text depression risk value is 0.8, text anxiety risk value is 0.9, text stress risk value is 0.9, and text crisis level risk value is 0.07; based on behavioral signal data and text signal data, the following features are calculated: speech rate feature is 5.6, average frowning intensity is 0.3, and fundamental frequency feature... The value is 97; the specific values ​​for each weighting coefficient are: The value is 0.25. The value is 0.25. The value is 0.15. The value can be 0.15 and... The value is 0.2; The value is 0.4. The value is 0.3; The value is 0.3; The value is 0.6. The value is 0.4; The value is 0.2; The value is 0.4; The value is 0.2; The value is 0.2.

[0059] In this embodiment, the preset threshold sequence is specifically as follows: When 0 ≤ R < 0.2, it is judged as a healthy level; when 0.2 ≤ R < 0.4, it is judged as a mild level. When 0.4 ≤ R < 0.6, it is judged as moderate level; when 0.6 ≤ R < 0.8, it is judged as severe level; when the total risk value is greater than or equal to 0.8, it is judged as crisis level.

[0060] The specific formula for calculating the depression risk value in this embodiment is as follows:

[0061] Therefore, the depression risk value in this embodiment is 0.45.

[0062] The specific formula for calculating the anxiety risk value in this embodiment is as follows: ; Therefore, the anxiety risk value in this embodiment is 0.66.

[0063] The specific formula for calculating the pressure load risk value in this embodiment is as follows: ; Therefore, the pressure load risk value in this embodiment is 0.62.

[0064] In this embodiment, the value of the crisis warning risk value directly depends on the text crisis level risk value, that is, the crisis warning risk value in this embodiment is 0.07.

[0065] The specific formula for calculating the total risk value in this embodiment is as follows: ; Therefore, the total risk value of this embodiment is 0.49.

[0066] Based on the specific value of the total risk score, the individual risk level can be determined as mild. The initial screening report is automatically generated as follows: "A has depression, anxiety and stress, and it is recommended to conduct AI interviews, provide training and companionship; there is no immediate crisis."

[0067] Reference Figure 3 The present invention proposes an AI-based psychological counseling system, comprising: The data acquisition module is used to collect physiological signal data, behavioral signal data, text signal data, and individual attribute data of an individual, including at least gender. The first processing module is used to extract physiological features from physiological signal data, behavioral features from behavioral signal data, and text features from text signal data. Based on the physiological features, behavioral features, and text features, and in combination with the individual attribute data, it calculates depression risk value, anxiety risk value, stress load risk value, and crisis warning risk value respectively to form a four-dimensional subject risk vector. The second processing module is used to calculate the total risk value based on the four-dimensional subject risk vector; The third processing module is used to compare the total risk value with a preset threshold sequence to determine the individual risk level, which includes health level, mild level, moderate level, severe level and crisis level. The data output module is used to automatically match consultation resources based on individual risk levels and generate scheduling plans and preliminary screening reports.

[0068] This embodiment also includes: a personalized AI interview module, which is responsible for simulating a senior consultant to conduct highly empathetic structured interviews and generate an evaluation report. Simultaneously, it performs real-time ethical compliance checks and quality assessments on the real-person consultation process, achieving closed-loop management of the entire process from individual assessment to service quality control. Specifically, this embodiment establishes a multi-dimensional knowledge archive mechanism for visitors, encompassing screening, companionship, coaching, interviews, and supervision, to achieve personalized diagnosis and process supervision linkage for each individual. It enables highly empathetic AI interviews and automatically generates structured reports; for real-person interviews, the system compares the transcripts of the real consultations with an ethical rules database in real time, generating sentence-by-sentence compliance suggestions and an overall score, providing interpretable supervisory conclusions for expert review. For visitors who have completed AI screening and are classified as L1-L2, the system automatically initiates AI interviews, synchronizing the screening results to the personalized AI interview module with a single click. The system also links to the visitor's AI companionship and coaching history in the background, allowing the personalized AI interview module to preview the visitor's situation (number of AI companionship sessions and chief complaint, AI coaching performance, AI interview history, total risk value, corresponding risk level, and sub-risk details) to achieve personalized diagnosis for each individual.

[0069] In this embodiment, it also includes: a companion and question-solving intelligent agent and a targeted push training module, which is used to build a continuous companionship and group intervention engine based on individual L0-L4 risk profiles and group four-dimensional risk statistics. It completes a closed-loop dialogue of "empathic reflection → interpretation clarification → attribution reconstruction → micro-goal setting → debriefing reminder" based on the visitor's real-time dialogue, realizing 24 / 7 high-empathy question answering and skills training; if high-risk expressions are detected, the safety handling process is immediately triggered, realizing a seamless switch from daily companionship to crisis interception.

[0070] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-based psychological counseling method, characterized in that, Includes the following steps: S1. Collect physiological signal data, behavioral signal data, text signal data, and individual attribute data of an individual, wherein the individual attribute data includes at least gender; S2. Extract physiological features from physiological signal data, behavioral features from behavioral signal data, and text features from text signal data. Based on physiological features, behavioral features, and text features, and in conjunction with the individual attribute data, calculate depression risk value, anxiety risk value, stress load risk value, and crisis warning risk value respectively to form a four-dimensional subject risk vector. S3. Calculate the total risk value based on the four-dimensional thematic risk vector; S4. Compare the total risk value with a preset threshold sequence to determine the individual risk level, which includes health level, mild level, moderate level, severe level and crisis level; S5. Automatically match consulting resources based on individual risk levels and generate scheduling plans and preliminary screening reports.

2. The AI-based psychological counseling method according to claim 1, characterized in that, The physiological signal data includes at least photoplethysmography (PPG) signals and blood oxygen saturation signals acquired by a finger-clip PPG sensor, and electroencephalogram (EEG) signals acquired by a wearable EEG sensor; the behavioral signal data includes at least speech signals acquired by a microphone and facial video signals acquired by a camera; the text signal data is obtained by converting the speech signals into text through speech recognition, or by directly receiving text data input by the user.

3. The AI-based psychological counseling method according to claim 2, characterized in that, The extraction of physiological features from physiological signal data specifically includes: The root mean square values ​​of heart rate and the difference between adjacent heartbeats are extracted from photoplethysmography (PPG) signals; the standard deviation of blood oxygen saturation is extracted from blood oxygen saturation signals; and the standard deviation of blood oxygen saturation is extracted from electroencephalogram (EEG) signals. Bandwidth power and The ratio of bandwidth power; The extraction of behavioral features from behavioral signal data specifically includes: Speech rate features and fundamental frequency features are extracted from speech signals, and the average intensity of frowning occurrences of facial frowning action units is extracted from facial video signals. The extraction of text features from text signal data specifically includes: By performing semantic analysis on the text using a large language model, the text can be directly output as a risk value for depression, anxiety, stress, and crisis.

4. The AI-based psychological counseling method according to claim 3, characterized in that, The depression risk value was obtained through the following method: based on gender information, a corresponding weighted calculation model was selected, and the text depression risk value, root mean square value of the difference between adjacent heartbeat intervals, fundamental frequency characteristics, standard deviation of blood oxygen saturation, and... Bandwidth power and The ratio of bandwidth power is weighted and the calculation result is limited to a predetermined range. The anxiety risk value is obtained by weighted summation of text anxiety risk value, speech rate characteristics, and average intensity of frowning, and the summation result is limited to a predetermined numerical range.

5. The AI-based psychological counseling method according to claim 3, characterized in that, The pressure load risk value is obtained by weighted summing of the text pressure risk value and the physiological pressure index calculated based on the root mean square value of the difference between adjacent heartbeat intervals, and limiting the summation result to a predetermined numerical range. The value of the crisis warning risk value is directly determined by the risk value of the text crisis level, and when a preset high-risk intent keyword is identified, the crisis warning risk value is set to the defined maximum value.

6. The AI-based psychological counseling method according to claim 3, characterized in that, The semantic analysis of text using a large language model specifically includes: The text obtained from speech recognition or the text data directly input by the user, along with the preset psychological analysis prompts, are input into the large language model. The large language model is used to perform semantic analysis on the text and directly output four independent text risk values ​​in parallel. The text risk values ​​include: text depression risk value, text anxiety risk value, text stress risk value, and text crisis level risk value. The large language model is configured to simultaneously evaluate and output a quantitative value of the risk level of the text in four psychological dimensions: depression, anxiety, stress, and crisis, based on the input text content. Each quantitative value is a value between 0 and 1.

7. The AI-based psychological counseling method according to claim 1, characterized in that, Step S3 specifically includes: The four-dimensional subject risk vector, composed of the depression risk value, anxiety risk value, stress load risk value and crisis warning risk value, is multiplied by a preset weight vector, and the result of the dot product operation is the total risk value. The preset weight vector contains configurable weight coefficients corresponding to each risk value in the four-dimensional subject risk vector.

8. The AI-based psychological counseling method according to claim 1, characterized in that, Step S4 specifically includes: The total risk value is compared to a preset threshold sequence consisting of five increasing thresholds; Among them, when the total risk value falls between the first threshold and the second threshold, it is judged as a healthy level; When the total risk value falls between the second and third thresholds, it is classified as a mild level. When the total risk value falls between the third and fourth thresholds, it is classified as a moderate level. When the total risk value falls between the fourth and fifth thresholds, it is classified as severe. When the total risk value is greater than or equal to the fifth threshold, it is judged to be at the crisis level.

9. The AI-based psychological counseling method according to claim 1, characterized in that, Step S5 specifically includes: Based on the preset treatment mapping rules, the corresponding type of consultation resources is matched to the individual risk level; The handling mapping rules include at least the following: mapping health level to targeted training resources, mapping mild and moderate levels to AI interview resources, mapping severe levels to human consultation resources with additional AI supervision resources, and mapping crisis level to human consultation resources and initiating crisis handling procedures. Based on the matched consultation resource type, query the corresponding resource pool status, and generate a scheduling plan containing specific time, consultant ID and contact channel based on the priority scheduling algorithm; Meanwhile, based on individual risk levels, total risk values, and four-dimensional thematic risk vectors, a preliminary screening report is generated through a large model, which includes a summary of objective indicators, an interpretation of risk factors, and suggested handling paths. For severe and crisis levels, the initial screening report will simultaneously generate a work order for manual review.

10. The AI-based psychological counseling method according to claim 1, characterized in that, Also includes: S6. Based on individual risk levels and initial screening reports, conduct highly empathetic structured interviews to generate transcripts of the interviews, key points of the interviews, and structured interview reports.

11. An AI-powered psychological counseling system, characterized in that, include: The data acquisition module is used to collect physiological signal data, behavioral signal data, text signal data, and individual attribute data of an individual, wherein the individual attribute data includes at least gender; The first processing module is used to extract physiological features from physiological signal data, behavioral features from behavioral signal data, and text features from text signal data. Based on the physiological features, behavioral features, and text features, and in combination with the individual attribute data, it calculates depression risk value, anxiety risk value, stress load risk value, and crisis warning risk value respectively to form a four-dimensional subject risk vector. The second processing module is used to calculate the total risk value based on the four-dimensional subject risk vector; The third processing module is used to compare the total risk value with a preset threshold sequence to determine the individual risk level, which includes a health level, a mild level, a moderate level, a severe level, and a crisis level. The data output module is used to automatically match consultation resources based on individual risk levels and generate scheduling plans and preliminary screening reports.

Images