A multi-scale psychological crisis early warning method and system

Abstract

This invention relates to the field of mental health assessment technology and discloses a multi-scale mental health crisis early warning method and system. The multi-scale mental health crisis early warning method includes: constructing a cross-scale knowledge graph and training a graph attention network, while simultaneously collecting high-dimensional fingerprints of respondents' answers and interactive behaviors to build a multimodal assessment dataset; using a temporal convolutional network and information entropy calculation to extract cognitive load shift features and output a behavioral fluctuation entropy feature vector; using a graph attention network to perform cross-scale semantic consistency analysis and generate a logical conflict detection matrix; outputting a corrected crisis assessment result through a dual-path parallel feature extraction network and a self-attention fusion mechanism; and finally generating a deep early warning report containing a risk profile and intervention recommendations. This invention can effectively identify respondents' concealment behavior in psychological assessments, improving the accuracy and foresight of mental health crisis early warning.

Description

Technical Field

[0001] This invention relates to the field of mental health assessment technology, and more specifically, to a multi-scale mental crisis early warning method and system. Background Technology

[0002] Psychological crisis early warning is an important component of mental health work in universities. Current technologies typically rely on standardized psychological scales to assess respondents, deriving risk levels by weighting and summing scores from various dimensions. However, this approach has significant limitations in practical application: scale assessments depend entirely on the respondents' subjective self-reports, making it impossible to verify the authenticity of their responses.

[0003] In real-world assessment scenarios, some respondents, concerned about privacy breaches, scholarship evaluations, or social judgment, consciously conceal their true psychological state and deliberately choose options that lean towards health. Traditional methods, relying solely on scores from a single scale, fail to capture the behavioral dynamics of respondents during the answering process, nor can they utilize the semantic logical connections between items across scales to verify the consistency of responses. This results in high-risk individuals exhibiting concealment behavior being missed.

[0004] Therefore, existing technologies have technical problems such as being unable to identify respondents' concealment behavior and insufficient accuracy in psychological crisis early warning. There is an urgent need for an early warning method that can detect concealment behavior from both the behavioral dynamics layer and the semantic logic layer and correct the evaluation results. Summary of the Invention

[0005] This invention provides a multi-scale psychological crisis early warning method and system, which solves the technical problems of being unable to identify respondents' concealment behavior and insufficient accuracy of psychological crisis early warning in related technologies.

[0006] This invention provides a multi-scale psychological crisis early warning method, comprising the following steps: S1. Obtain psychological assessment task requests, construct a cross-scale knowledge graph and train it through a graph attention network, while simultaneously collecting high-dimensional fingerprints of respondents' answers and interactive behaviors in real time to obtain a cross-scale knowledge graph and a multimodal assessment dataset containing the original assessment results; the cross-scale knowledge graph is constructed by identifying question pairs and labeling logical association types through semantic similarity calculation, with questions as nodes and logical association relationships as edges. S2, acquire a multimodal assessment dataset, and extract cognitive load shift features using a combination of temporal convolutional networks and information entropy calculation, outputting a behavioral fluctuation entropy feature vector and a convolutional feature sequence; the behavioral fluctuation entropy feature vector is constructed by concatenating the behavioral depth feature vector extracted by the temporal convolutional network with the behavioral entropy value statistical feature vector calculated based on the Shannon entropy formula; S3. Based on the knowledge graph and the answer content, a graph attention network is used to perform cross-scale semantic consistency analysis, generate a logical conflict detection matrix and output a logical conflict feature vector; the logical conflict feature vector determines whether there is a logical deviation in the item pair based on the logical association type and a preset consistency threshold, and constructs the number and severity of deviations on each psychological dimension. S4 receives the behavioral fluctuation entropy feature vector and the logical conflict feature vector, and uses a dual-path parallel feature extraction network and a self-attention fusion mechanism to perform a comprehensive evaluation, and outputs the corrected crisis assessment result. S5, based on the revised crisis assessment results, uses risk level determination rules and dynamic evolution prediction strategies to generate an in-depth early warning report containing risk profiles and intervention recommendations.

[0007] In a preferred embodiment, S1 includes: A pre-trained language model was used to semantically vectorize the text of all psychological scale items. Semantic similarity calculation was used to identify semantically related item pairs. Experts in the field of clinical psychology reviewed and labeled the item pairs to determine the logical relationship type between them. The logical relationship type includes logical support relationship and logical conflict relationship, and the association strength weight was labeled. The labeled logical relationships are constructed into a cross-scale knowledge graph, where nodes represent questions and edges represent logical relationships between questions. The training of the graph attention network includes: constructing a graph attention network on the cross-scale knowledge graph, using a multi-head attention mechanism to calculate the attention weights between nodes and their neighbors, and using a contrastive learning framework to train the graph attention network with historical test data, so that the graph attention network learns question representations that are sensitive to the consistency of individual response patterns.

[0008] In a preferred embodiment, S1 further includes: When respondents take the test, physical layer features, dynamic layer features and contextual features are collected to form a high-dimensional fingerprint of the interaction behavior. The physical layer features include answer time, visual processing time, hesitation trajectory distance and backtracking correction rate. The dynamic layer features include the coordinate sequence of option switching trajectory and mouse movement trajectory, as well as speed and acceleration changes. The collected raw behavioral data were normalized to generate a behavioral fingerprint baseline for each respondent. The answer content data and interaction behavior data were correlated and integrated to construct a multimodal assessment dataset. Each record in the dataset contains respondent identifier, question identifier, option score, question type and multidimensional high-dimensional fingerprint features. The question types include crisis sensitivity type and routine assessment type.

[0009] In a preferred embodiment, S2 includes: The response time, first response delay, number of option switching times, and hesitation trajectory length of each question were extracted from the multimodal assessment dataset. Based on the question type, the questions were divided into regular assessment questions and crisis-sensitive questions. We extracted the behavioral characteristics of respondents on routine assessment questions, and calculated the mean and standard deviation of answer time, first response delay, number of option switching and hesitation trajectory length to establish an individualized behavioral fingerprint baseline. For each crisis-sensitive question, we calculated the standardized deviation score of its behavioral characteristics relative to the individualized behavioral fingerprint baseline, which was defined as the cognitive load increment.

[0010] In a preferred embodiment, S2 further includes: The response time sequence, response delay sequence, switching number sequence, and trajectory length sequence of respondents are used as multi-channel inputs. A one-dimensional temporal convolutional network with multi-scale convolutional kernels is constructed to extract deep features. The behavior depth feature vector is obtained by global max pooling, while the convolutional feature sequence is preserved. The response time of respondents on crisis-sensitive questions is discretized into intervals, a discrete probability density function is constructed, and the duration entropy is calculated based on the Shannon entropy formula. The entropy values ​​of the number of option switching and the first response delay are calculated separately to obtain the comprehensive behavioral entropy value. The behavioral depth feature vector is concatenated with the statistical feature vector containing the comprehensive behavioral entropy value, duration entropy, switching entropy, delay entropy and deviation mean to construct the behavioral fluctuation entropy feature vector.

[0011] In a preferred embodiment, S3 includes: For each completed question, the question text and the text of the options selected by the respondent are concatenated to form a combined text, which is then input into a pre-trained language model to obtain the response content vector. The response content vector is used as the dynamic initial feature of the corresponding node in the cross-scale knowledge graph, and the feature is propagated and updated through a graph attention network to obtain the response context representation vector for each question. Iterate through all logically related question pairs, calculate the attention-weighted similarity between the response context representation vectors, and determine logical deviations when the similarity is lower than the preset positive consistency threshold for logically supported question pairs, and logical deviations when the similarity is higher than the preset negative consistency threshold for logically conflicting question pairs. Calculate the cumulative value of the number of logical deviations and the severity of deviations in each psychological dimension, calculate the global conflict density, and construct a logical conflict feature vector.

[0012] In a preferred embodiment, S4 includes: The dual-path parallel feature extraction network includes a behavioral dynamics branch and a semantic conflict branch. The behavioral dynamics branch receives the convolutional feature sequence, uses a long short-term memory network to extract long-range temporal feature representations, and then performs feature transformation through a fully connected neural network to output a high-level feature vector of the behavioral layer. The semantic conflict branch receives the logical conflict feature vector, constructs a psychological dimension association graph by using the conflict intensity of each psychological dimension as node features, performs feature aggregation using a graph attention network, and then performs feature transformation through a fully connected neural network to output the high-level feature vector of the logical layer. The self-attention fusion mechanism includes two gating units: a behavioral feature gate and a logical feature gate. The high-level feature vectors of the behavioral layer and the high-level feature vectors of the logical layer are concatenated and then input into the corresponding fully connected layers to obtain the gating weight vectors. Each high-level feature vector is multiplied element-wise with the corresponding gating weight vector and then added element-wise to obtain the fused feature vector.

[0013] In a preferred embodiment, S4 further includes: The fused feature vector is mapped to a concealment score through a fully connected network. The concealment score ranges from 0 to 1. When the concealment score exceeds a preset concealment judgment threshold, it is determined that the respondent has concealed behavior. Different correction strategies are adopted according to the scoring direction attribute of the psychological dimension. For negative scoring dimensions, the correction coefficient is set to be greater than 1 to adjust the original score, and for positive scoring dimensions, the correction coefficient is set to be less than 1 to adjust the original score. The original scale's dimension scores are multiplied by their corresponding correction coefficients to obtain the corrected dimension scores. The overall crisis score is then recalculated based on these corrected dimension scores.

[0014] In a preferred embodiment, S5 includes: A four-level risk rating system is pre-set, which divides the range of the comprehensive crisis score into four intervals. The score is determined to be extremely high risk if it exceeds the first pre-set threshold, high risk if it is between the second pre-set threshold and the first pre-set threshold, medium risk if it is between the third pre-set threshold and the second pre-set threshold, and at least at concern level if it is below the third pre-set threshold. The dynamic evolution prediction strategy includes: extracting feature vectors from respondents' historical assessments, calculating the Euclidean distance between the current assessment feature vector and the historical assessment feature vector in the feature space as the magnitude of psychological state change, fitting the time evolution trajectory to calculate the trajectory slope as the speed of psychological state change, and triggering a dynamic early warning mechanism when the speed of change exceeds a preset deterioration speed threshold; the risk profile includes explanations of concealment, core symptom localization, and output of targeted intervention strategies.

[0015] In a preferred embodiment, a multi-scale psychological crisis early warning system is used to perform the steps in the above-described multi-scale psychological crisis early warning method, including: The graph construction and acquisition module is used to obtain psychological assessment task requests, construct a cross-scale knowledge graph and train it through a graph attention network, and simultaneously collect high-dimensional fingerprints of respondents' answers and interaction behaviors in real time to obtain a cross-scale knowledge graph and a multimodal assessment dataset containing the original assessment results. The cross-scale knowledge graph is constructed by identifying question pairs and labeling logical association types through semantic similarity calculation, with questions as nodes and logical association relationships as edges. The cognitive load shift feature extraction module is used to acquire a multimodal assessment dataset. It uses a combination of temporal convolutional networks and information entropy calculation to extract cognitive load shift features and outputs a behavioral fluctuation entropy feature vector and a convolutional feature sequence. The behavioral fluctuation entropy feature vector is constructed by concatenating the behavioral depth feature vector extracted by the temporal convolutional network with the behavioral entropy value statistical feature vector calculated based on the Shannon entropy formula. The cross-scale semantic consistency analysis module is used to perform cross-scale semantic consistency analysis based on the knowledge graph and the answer content using a graph attention network, generate a logical conflict detection matrix and output a logical conflict feature vector; the logical conflict feature vector determines whether there is a logical deviation between the items based on the logical association type and a preset consistency threshold, and constructs the system by counting the number and severity of deviations on each psychological dimension; The dual-path feature fusion and comprehensive evaluation module receives behavioral fluctuation entropy feature vectors and logical conflict feature vectors, and uses a dual-path parallel feature extraction network and self-attention fusion mechanism to perform comprehensive evaluation, outputting a corrected crisis assessment result. The in-depth early warning report generation module is used to generate in-depth early warning reports containing risk profiles and intervention recommendations based on the revised crisis assessment results, risk level determination rules, and dynamic evolution prediction strategies.

[0016] The beneficial effects of this invention are as follows: By constructing a cross-scale semantic association knowledge graph and combining it with a graph attention network, we can conduct cross-scale semantic consistency analysis on the responses of respondents across multiple scales. By detecting logical deviations in responses from the semantic logic level, we can identify the concealment strategies of respondents who deliberately conceal in the core crisis dimension and expose contradictions in the peripheral dimension, thus making up for the inability of traditional scale assessment to verify the authenticity of responses. By collecting high-dimensional fingerprints of respondents' interactive behaviors during the question-and-answer process, and using temporal convolutional networks and information entropy calculations to extract cognitive load shift features from the behavioral dynamics layer, the behavioral layer features and logical layer features are adaptively fused through a dual-path parallel feature extraction network and a self-attention fusion mechanism. The resulting masking score is then used to correct the original scale assessment results, thereby improving the ability of psychological crisis early warning to detect highly masked respondents. Attached Figure Description

[0017] Figure 1 This is a flowchart of a multi-scale psychological crisis early warning method according to the present invention; Figure 2 This is a flowchart of a multi-scale psychological crisis early warning method according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a multi-scale psychological crisis early warning method, such as... Figures 1 to 2 As shown, it includes the following steps: S1. Obtain psychological assessment task requests, construct a cross-scale knowledge graph and train it through a graph attention network, while simultaneously collecting high-dimensional fingerprints of respondents' answers and interactive behaviors in real time to obtain a cross-scale knowledge graph and a multimodal assessment dataset containing the original assessment results; the cross-scale knowledge graph is constructed by identifying question pairs and labeling logical association types through semantic similarity calculation, with questions as nodes and logical association relationships as edges. This step aims to establish the data foundation and logical support for the entire early warning system. Traditional early warning methods rely solely on the simple summation of scores from a single scale, failing to capture the deep psychological logical connections between items on different scales and struggling to characterize the micro-behavioral patterns of respondents during the answering process. This step lays the foundation for subsequent logical conflict detection and cognitive load analysis by constructing a cross-scale semantic association knowledge graph and a high-dimensional fingerprint feature library of interactive behaviors. This step includes an offline construction phase and an online data collection phase.

[0020] The offline construction phase, completed before the formal system deployment, builds a cross-scale semantic association knowledge graph. The system first deconstructs all integrated psychological scales, which include not only structured single-choice questions but also open-ended text statements. The system uses natural language processing (NLP) technology to semantically vectorize all items. Specifically, the system uses a pre-trained language model based on deep semantic understanding to encode the question text, preferably employing a BERT model fine-tuned for the Chinese psychological domain. The text description of each question is encoded into a semantic vector, with a default dimension of 768. Through semantic vector similarity calculation, semantically related question pairs are initially identified.

[0021] Then, experts in the field of clinical psychology manually reviewed and labeled the initially identified item pairs. Based on clinical diagnostic criteria such as the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) and the International Classification of Diseases, Eleventh Revision (ICD-11), the experts determined whether there was a psychological logical connection between the item pairs. Logical connections were categorized into two types. The first type is a supporting logical relationship, meaning that the psychological symptoms or states assessed by the two items usually co-occur or reinforce each other clinically. For example, if a respondent admits to feeling that life is meaningless in scale A, they should also have a negative evaluation of being hopeful for the future in scale B; the two should show positive consistency. The second type is a conflicting logical relationship, meaning that the psychological symptoms or states assessed by the two items are clinically mutually exclusive or contradictory. For example, there is a conflicting logical relationship between frequent insomnia symptoms and good sleep quality; the two should show negative opposition. The experts labeled each pair of logically related items with the type of relationship and the weight of the relationship strength.

[0022] The system constructs a knowledge graph from the logical relationships annotated by experts. Each node in the graph represents a question, with attributes including the question text, the scale it belongs to, the assessment dimension, core symptom tags, and a semantic vector. The core symptom tags indicate the type of psychological symptom assessed by the question, including depressed mood, negative thoughts, sleep disorders, and social withdrawal. Edges in the graph represent logical relationships between questions, with attributes including relationship type and relationship strength. The system pre-defines two key edge attributes: logically supporting edges to indicate consistent relationships between questions, and logically conflicting edges to capture completely contradictory statements. Based on psychological diagnostic criteria, the system pre-defines probability weights for logical divergence between different question pairs, providing a benchmark for subsequent logical consistency analysis. This knowledge graph is stored in a graph database, supporting efficient relational queries.

[0023] To enhance the knowledge graph's ability to model deep semantic relationships between items, the system constructs a graph attention network on the knowledge graph. The motivation for designing the graph attention network is that there are complex multi-level relationships between items in psychological scales. Simple cosine similarity calculation cannot capture such multi-hop semantic dependencies. By propagating information on the graph structure, the graph attention network allows each item node to aggregate the semantic information of its neighboring nodes, thereby learning a item representation that includes contextual relationships.

[0024] The graph attention network employs a two-layer stacked structure. The first graph attention layer receives the initial semantic vector of the question as input. For each node in the graph, it calculates the attention weights between it and all its neighboring nodes. The calculation of these attention weights uses a multi-head attention mechanism, with a default value of 8 heads. Each attention head independently calculates the relevance score between node pairs, and then the results are normalized using softmax to obtain the attention weights. Nodes then weight and aggregate the features of their neighboring nodes using these attention weights to obtain an updated node representation. The default output dimension of the first layer is preferably 256 dimensions. The second graph attention layer further propagates information based on the first layer, also using a multi-head attention mechanism, with a default output dimension of preferably 128 dimensions. After these two layers of propagation, the representation vector of each question node not only contains its own semantic information but also incorporates contextual information from first-order and second-order neighbors, forming a richer contextual representation of the question.

[0025] The training of the graph attention network employs a contrastive learning framework. Utilizing extensive historical assessment data, the system treats the responses of the same respondent completed at different times as positive sample pairs, constructing a training objective for contrastive learning. The system employs a filtering mechanism to exclude sample pairs where the psychological state changes between assessments. Specifically, records where the original scale score changes by more than a preset threshold, or records where subsequent psychological interviews confirm significant life events experienced during that period, are removed from the positive sample pairs, ensuring that positive sample pairs reflect relatively stable psychological characteristic patterns. By maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs, the graph attention network learns item representations sensitive to the consistency of individual response patterns. This training process does not require manual annotation of masked behaviors; self-supervised learning is achieved using the natural pairing relationships of historical assessment data.

[0026] For initial deployments or scenarios with insufficient historical data, the system supports the following initialization strategies: First, pre-training is performed using publicly available psychological assessment datasets or anonymized data shared by other institutions, with the parameters of the pre-trained model used as initial values ​​through transfer learning. Second, rule-based knowledge graph initialization is employed, using only expert-annotated logical relationships. The graph attention network can be trained after accumulating a certain amount of data during system operation; before that, the system uses the initial semantic vectors of the questions for logical conflict detection. Third, pre-trained graph neural networks from general domains are used for parameter initialization, allowing for fine-tuning with a small amount of local data to quickly adapt to specific application scenarios. During actual operation, the system continuously accumulates data and periodically performs incremental training or retraining of the graph attention network to achieve continuous model optimization.

[0027] To further enhance the completeness of the knowledge graph, the system utilizes historical evaluation data for statistical verification and supplementation. For question pairs with frequently co-occurring abnormal response patterns in historical data, the system submits them as candidate associations for secondary annotation by experts, even if no pre-defined associations have been established. New associations confirmed by experts are dynamically added to the knowledge graph, enabling continuous optimization of the knowledge graph.

[0028] The online data collection phase is executed during real-time assessment by respondents, constructing a high-dimensional fingerprint feature database of interactive behaviors. The system receives assessment task requests from the psychological assessment platform, which include the respondent's unique identifier and a list of scales to be completed in this assessment. The system retrieves corresponding questions from the question bank based on the scale list and generates an assessment session according to a preset presentation order. The assessment session is presented to the respondent through a web frontend or mobile application, and the respondent answers the questions on their terminal device.

[0029] When respondents take the assessment via their terminal devices, the backend engine captures their behavioral data throughout the entire process with a sampling precision of 10 milliseconds. The system captures the respondents' interactive operations through the front-end monitoring module, collecting physical layer features, dynamic layer features, and contextual features to form a high-dimensional fingerprint of the interactive behavior.

[0030] Physical layer features include: the question presentation time, i.e., the timestamp when the question content is fully loaded and displayed on the screen; the visual processing time from the completion of question loading to the first cursor movement, defined as the time difference between the first interaction and the question presentation time, which reflects the speed at which respondents process visual information; the hesitation trajectory distance between options, defined as the total path length of the mouse or touch point moving between different option areas, which reflects the degree to which respondents hesitate between options; and the backtracking correction rate before clicking confirm, defined as the ratio of the number of times respondents modify options before final submission to the total number of times they answer questions, which reflects the stability of respondents' decisions.

[0031] The dynamics layer features include: option switching trajectory, which is the complete sequence of respondents moving and clicking between different options, including the moment and duration of each mouse hover over an option; the coordinate sequence of the mouse movement trajectory and its speed and acceleration changes; if the hardware supports it, the system also analyzes the key pressure distribution or touch area changes when users select questions with different sensitivity levels.

[0032] Contextual features include: recording the consistency of users' responses when encountering logically related questions, that is, the similarity of respondents' answering time patterns and interaction trajectory patterns for question pairs marked as logically supporting or logically conflicting relationships in the knowledge graph.

[0033] The system normalizes the collected raw behavioral data. It calculates the answering time for each question, defined as the time difference between the final submission and the question presentation. It also calculates the initial response latency, i.e., the visual processing time. Finally, it calculates the number of option switches, defined as the total number of times the respondent switches between different options. By normalizing these features, the system generates a dynamic behavioral fingerprint baseline for each respondent, used to filter noise caused by device performance fluctuations or differences in individual user writing and reading habits.

[0034] The system integrates and correlates answer content data and interaction behavior data to construct a multimodal assessment dataset. Each record in the dataset corresponds to one question and includes multi-dimensional high-dimensional fingerprint features such as respondent identifier, question identifier, scale identifier, option score, question type, answer duration, initial response delay, number of option switching times, hesitation trajectory length, mouse movement trajectory coordinates, and touch pressure. The question type field is categorized as either crisis-sensitive or standard assessment. This classification label was pre-labeled by psychology experts during the system deployment phase and stored in the question bank along with the question content for subsequent behavioral analysis steps. Crisis-sensitive questions directly relate to core crisis symptoms, such as depressive mood, negative thoughts, severe anxiety, and self-harm behavior. Standard assessment questions relate to general psychological states, personality traits, or life events. This multimodal assessment dataset serves as a unified input for subsequent analysis steps.

[0035] The system calculates the raw assessment results based on the collected responses, following the standard scoring methods for each psychological scale. Specifically, for each psychological dimension, the system weights and sums the scores of all options under that dimension according to the scale's prescribed weights to obtain the raw score for that dimension. For scales containing reverse-scored items, the system first converts the scores of the reverse items, transforming the raw option scores into reverse scores using a formula: the reverse score equals the scale's maximum score plus the minimum score minus the raw score, and then sums them. The system compares the raw scores of each dimension with the scale norms to obtain standardized scores. The comprehensive crisis score is calculated based on a weighted combination of scores from multiple dimensions or a decision rule. The weighting coefficients are set according to the contribution of each dimension to crisis early warning as determined by clinical research; for example, the depression and negative ideation dimensions typically have higher weights. This raw assessment result includes the scores of each dimension, the comprehensive crisis score, and the preliminary risk level, serving as a baseline for traditional scale assessments and for subsequent masking degree correction steps.

[0036] S2, acquire a multimodal assessment dataset, and extract cognitive load shift features using a combination of temporal convolutional networks and information entropy calculation, outputting a behavioral fluctuation entropy feature vector and a convolutional feature sequence; the behavioral fluctuation entropy feature vector is constructed by concatenating the behavioral depth feature vector extracted by the temporal convolutional network with the behavioral entropy value statistical feature vector calculated based on the Shannon entropy formula; The core of this step lies in leveraging the scientific principle that concealing behavior requires cognitive resources to transform behavioral fluctuations into quantifiable masquerade features. Based on the dual-process theory of cognitive psychology, when respondents attempt to conceal their true psychological state, their brains operate in a dual-process state: one path is the feedback of the true psychological state, and the other is the logic of the disguised false response. This dual-process involves multiple stages: first, the automatic activation of the true psychological state; second, the conscious inhibition of the true response; third, the construction of a false response that conforms to the desired image; and finally, the evaluation and confirmation of the rationality of the false response. This multi-stage processing requires more cognitive resources than the single-stage processing of honest responses, leading to an increased cognitive load. This increased cognitive load manifests at the behavioral level as irregular fluctuations in the rhythm of answering questions, specifically as abnormally long or short answering times, unstable response delays, and increased option switching behavior. This dual-process inevitably leads to microscopic disturbances in the behavioral rhythm. This step uses a mathematical model to extract abnormal shifts in cognitive load, identifying local abnormal patterns in the behavioral sequence.

[0037] The system extracts interaction behavior data from the multimodal assessment dataset constructed by S1, including the answering time for each question, initial response delay, number of option switching times, and hesitation trajectory length. The system categorizes questions into general assessment questions and crisis-sensitive questions based on the question type field. The categorization criteria are based on the psychological dimensions and symptom severity assessed by the questions. Crisis-sensitive questions directly address core crisis symptoms, including depressive mood, negative thoughts, severe anxiety, and self-harming behavior. General assessment questions cover general psychological states, personality traits, or life events.

[0038] The system first establishes an individualized behavioral fingerprint baseline for the respondents. The system categorizes assessment questions into low-sensitivity background questions and high-sensitivity crisis questions. It extracts the respondent's behavioral characteristics across all standard assessment questions, calculating the mean and standard deviation of response time, denoted as the baseline mean and standard deviation. It also calculates the mean and standard deviation of initial response delay, denoted as the baseline mean and standard deviation. Furthermore, it calculates the mean and standard deviation of the number of option switching attempts, denoted as the baseline mean and standard deviation. Finally, it calculates the mean and standard deviation of the hesitation trajectory length, denoted as the baseline mean and standard deviation. These statistics characterize the respondent's natural answering habits and rhythm under no psychological defense pressure, serving as an individualized reference benchmark for subsequent anomaly detection, filtering out noise caused by device performance fluctuations or differences in individual user habits.

[0039] The system then extracts the cognitive load increment of respondents on highly sensitive crisis questions. For each crisis-sensitive question, the system extracts the response time, initial response delay, number of option switching attempts, and hesitation trajectory length. The system defines the cognitive load increment by calculating the degree of deviation of the respondent's response characteristics on highly sensitive questions relative to their baseline for background questions. Specifically, the deviation in response time is defined as the response time for that question minus the mean of the baseline response time, divided by the standard deviation of the baseline response time, to obtain a standardized deviation score. The deviation in initial response delay and the deviation in the number of option switching attempts are calculated using the same standardization method. Since the hesitation trajectory length may be close to zero on regular questions, the system uses a logarithmically transformed relative ratio method to calculate the deviation. These deviations quantify the degree of behavioral abnormality of respondents on crisis-sensitive questions relative to their personal norms.

[0040] The system constructs a one-dimensional temporal convolutional network to extract deep features from respondents' answer behavior sequences. Masking behavior exhibits locality in the temporal dimension, meaning that abnormal behavior tends to erupt in concentrated bursts on a few consecutive crisis-sensitive questions. The local receptive field characteristic of convolutional neural networks is naturally well-suited for capturing this local temporal pattern. The system uses the respondents' answer duration sequence, response delay sequence, number of switching sequences, and trajectory length sequence as inputs to four channels, respectively, to construct a multi-channel temporal signal.

[0041] The temporal convolutional network employs a multi-scale convolutional kernel design. The first convolutional layer contains three sets of parallel convolutional operations, with kernel sizes preferably using a combination of 3, 5, and 7, corresponding to short-term fluctuations, medium-term trends, and long-term patterns, respectively. The default number of kernels in each set is preferably 64. The convolutional operations adopt a causal convolutional pattern, ensuring that feature extraction at the current time step depends only on historical information, which aligns with practical application scenarios. A ReLU activation function is applied after the convolutional layer to introduce non-linearity. The second convolutional layer further extracts high-level abstract features based on the first layer, with a kernel size preferably of 3 and a default number of kernels preferably of 128. After two convolutional layers, the system obtains a convolutional feature sequence that still retains its temporal structure. Each time step of this sequence corresponds to an abstract representation of the behavioral features of a question, with a dimension of 128. The system extracts the most obvious abnormal features from the entire temporal sequence through global max pooling, obtaining a fixed-dimensional behavioral depth feature vector with a dimension of 128. The system also retains the convolutional feature sequence as an intermediate output, which preserves the temporal dependencies of the question-answering process for subsequent steps to perform long-term temporal dependency analysis.

[0042] The temporal convolutional network is trained using a supervised learning framework. The system constructs a training set using a small number of expert-annotated masked and honest samples. Expert annotations are based on the results of subsequent psychological interviews. For respondents who admitted to masking their performance during the assessment, their evaluation records are labeled as masked samples. For respondents who showed high cooperation and consistency during the interview, their evaluation records are labeled as honest samples. The default number of labeled samples is 200, with half masked samples and half honest samples. The convolutional network is trained by minimizing the binary cross-entropy loss function, using the Adam optimizer with a default learning rate of 0.001. To prevent overfitting due to small sample training, an early stopping strategy is employed, terminating training when the validation set loss no longer decreases for several consecutive epochs. L2 regularization is introduced to constrain the norm of the network parameters, with a default regularization coefficient of 0.0001. A Dropout layer is added after the convolutional layers, with a default Dropout ratio of 0.3, randomly dropping some neuron activations to enhance the model's generalization ability. The training process employs data augmentation strategies, such as time stretching and adding Gaussian noise to the behavioral sequences, to expand the diversity of training samples and improve the robustness of the model.

[0043] For scenarios with insufficient training data, the system supports a few-shot learning strategy. Through a meta-learning framework, the system can quickly adapt to new scenarios with a small number of labeled samples. The system also supports transfer learning, using temporal convolutional network models trained in other psychological assessment scenarios for parameter initialization, and can be deployed and used after fine-tuning with a small amount of local data. In addition, the system can adopt a semi-supervised learning method, using a large amount of unlabeled assessment data for self-supervised pre-training to learn a general representation of behavioral sequences, and then using a small amount of labeled data for supervised fine-tuning.

[0044] To enhance the interpretability of the model and fully leverage the advantages of statistical methods, the system supplements the deep features extracted by the convolutional network with statistical features based on information entropy. This invention introduces the concept of information entropy to quantify the degree of disorder in behavioral sequences. The system treats the respondents' answer time series and interaction trajectory sequences as random signals, constructs a response probability distribution function based on the respondents' performance on questions of similar difficulty, and calculates the disorder of this distribution using the Shannon entropy formula.

[0045] To address the characteristics of response time, the system discretizes the time values ​​into intervals. Based on the time distribution of respondents on regular questions, the time axis is divided into several intervals, with the default width of each interval set to half the standard deviation of the time baseline. The frequency of occurrence of each time interval for crisis-sensitive questions is statistically analyzed to construct a discrete probability density function for the time distribution. Shannon entropy is then calculated based on this probability density function. The Shannon entropy is calculated as follows: First, for each time interval, the proportion of questions within that interval to the total number of crisis-sensitive questions is calculated as the probability value for that interval; then, the negative logarithm of the probability value for each interval is calculated; finally, the probability values ​​of all intervals are multiplied by their negative logarithms and summed to obtain the time entropy value. If the time distribution of respondents on crisis-sensitive questions is highly concentrated, it indicates that their response rhythm is stable and the entropy value is low. If the time distribution is scattered and irregular, it indicates that the response rhythm is chaotic and the entropy value is high. If the entropy value of a respondent's behavioral sequence suddenly increases when facing crisis-sensitive questions, it indicates that they are experiencing intense conflict in their response choices.

[0046] The system performs similar entropy calculations on the frequency of option switching. It statistically analyzes the frequency distribution of respondents switching options zero, once, twice, three times, or more on crisis-sensitive questions, and calculates the entropy value of the switching behavior. For the first response delay feature, it also performs interval discretization and entropy calculation. The system combines the three indicators—duration entropy, switching entropy, and delay entropy—to obtain the respondent's comprehensive behavioral entropy value on crisis-sensitive questions.

[0047] The system monitors the evolution of entropy values ​​in real time as the test progresses, identifying patterns in the evolution of answering strategies. If the entropy value shifts from a stable state in the early stages of the assessment to a periodic oscillation in the later stages, it indicates that the respondent may have entered a defensive response mode, consciously selecting questions and disguising their responses. The system records characteristic parameters of the entropy value evolution trajectory, including the mean, variance, maximum, minimum, and peak value of the entropy value, as well as the peak value of the rate of change of entropy.

[0048] The system concatenates the deep feature vector extracted by the temporal convolutional network with the statistical feature vector calculated based on information entropy to construct a behavioral fluctuation entropy feature vector. The feature vector includes: 128-dimensional deep features output by the convolutional network, comprehensive behavioral entropy value, duration entropy, switching entropy, delay entropy, mean duration deviation, mean switching deviation, mean delay deviation, and entropy fluctuation variance, among other statistical features. The total dimension of the feature vector is approximately 130 to 140 dimensions, with the specific dimension depending on the type and quantity of collected behavioral features. For example, if the hardware supports touch pressure detection, pressure-related entropy and deviation features will be added. Deep features capture complex nonlinear patterns and local temporal dependencies in the behavioral sequence, while statistical features provide interpretable anomaly measurements; the two complement each other. The behavioral fluctuation entropy feature vector, as a quantitative representation of the respondents' behavioral layer masking features, is output to subsequent fusion steps. Simultaneously, the system retains the intermediate output of the temporal convolutional network, namely the convolutional feature sequence. This sequence has a dimension equal to the number of questions multiplied by 128, preserving complete temporal structure information for long-range temporal dependency analysis in subsequent fusion steps.

[0049] S3. Based on the knowledge graph and the answer content, a graph attention network is used to perform cross-scale semantic consistency analysis, generate a logical conflict detection matrix and output a logical conflict feature vector; the logical conflict feature vector determines whether there is a logical deviation in the item pair based on the logical association type and a preset consistency threshold, and constructs the number and severity of deviations on each psychological dimension. This step addresses the technical challenge of identifying high-IQ deception by employing a logical verification method that uses the respondent's own arguments against them. Even if respondents can control their answering behavior, maintaining a perfect logical loop across a multi-scale question bank of hundreds of questions is difficult. This step utilizes a knowledge graph constructed with S1 to perform dynamic logical consistency checks, uncovering contradictions in respondents' responses at the semantic level.

[0050] The system calls the cross-scale semantic association knowledge graph and graph attention network already constructed in S1 to extract the current respondent's answer content data and construct a response semantic representation for each completed question. When the respondent completes a key item in scale A, the system uses the knowledge graph to index subsequent items with strong logical associations distributed in scales B and C in real time, and dynamically extracts implicit logical association pairs.

[0051] The specific method for constructing the semantic representation of responses is as follows: For each question, the system concatenates the question text with the text of the options selected by the respondent to form a question-option combination text. For example, for the question "I feel depressed" and the option "often," the concatenation is "I feel depressed, often." This combination text is then input into a pre-trained BERT model for encoding, and the sentence vector representation of the last layer of the model is extracted to obtain a response content vector with a dimension of 768. This vector contains not only the semantic information of the question but also the semantic information of the respondent's specific options, enabling a more accurate representation of the respondent's actual response meaning on that question.

[0052] The system inputs the respondent's response vector into a graph attention network, where features are propagated and updated under the guidance of the knowledge graph structure. For each question completed by the respondent, the system locates the corresponding node in the knowledge graph, using the respondent's response vector as the dynamic initial feature of that node, replacing the static semantic vector stored in that node during the pre-deployment phase. The edge structure of the graph remains fixed during real-time inference; only the initial features of the nodes are dynamically updated based on the respondent's actual responses. Through two layers of propagation in the graph attention network, the features of each node are guided by the fixed edge structure of the graph, aggregating the response information of its logically related questions. For example, if question A and question B have a logical supporting relationship, the node features of question A will absorb the response information of question B through the attention mechanism, thus reflecting the degree of correlation between the responses of the two questions in the feature space. After processing by the graph attention network, the system obtains a response context representation vector for each question, with a dimension of 128. This vector integrates the response semantics of the question itself and the response information of its logically related questions.

[0053] The system iterates through all questions completed by the respondent. For each question, it queries the knowledge graph for other questions that have a logical relationship with it. If related questions have also been completed by the respondent, they form a question pair, which undergoes consistency checking to generate a conflict detection matrix. The system then converts the respondent's option scores into coordinates in a semantic vector space for vector conflict determination.

[0054] For logically supported question pairs, the system extracts the response context representation vectors of the two questions after processing by a graph attention network and calculates the attention-weighted similarity between the vectors. The specific calculation method is as follows: first, calculate the dot product of the two vectors; then, scale by dividing by the square root of the vector dimension; finally, map to the interval of negative one to positive one using a hyperbolic tangent activation function. If the responses to two questions are logically mutually supportive, such as life satisfaction and future expectations (questions that should present a positive support relationship), the respondents' options should be consistent, and their response context representation vectors should show a similar distribution in the feature space, with the attention-weighted similarity close to positive one. The system sets a positive consistency threshold, which is determined through a grid search on the validation dataset, with a default value of 0.6. If the cosine of the angle between the related question pairs deviates, and the attention-weighted similarity is lower than this threshold, it is judged as a logical deviation, and a conflict node is recorded in the conflict detection matrix.

[0055] For logically conflicting question pairs, the system also calculates the attention-weighted similarity of the response context representation vectors. If the responses to two questions are logically conflicting, their response context representation vectors should exhibit opposite distributions in the feature space, and the attention-weighted similarity should be close to negative one or at least less than zero. The system sets a negative consistency threshold, with a default value of 0.2, indicating that the two responses should not show a significant positive correlation. When the attention-weighted similarity is higher than this threshold, it is considered a logical deviation.

[0056] The system performs multidimensional deviation analysis. The matrix records not only the occurrence of conflict but also its depth. Each logical deviation event is recorded, including the involved item pair, the severity of the deviation, and the corresponding psychological dimension. The severity of the deviation is quantified by the deviation between the actual similarity and the expected threshold; a larger deviation indicates a more severe deviation. Psychological dimensions are categorized based on the symptom types assessed by the items, including dimensions of depression, anxiety, interpersonal relationships, and somatization.

[0057] For example, respondents demonstrated logical rigor in core crisis dimensions such as negative thoughts, but acknowledged symptoms in peripheral dimensions such as somatic symptoms, including frequent headaches and insomnia. This mismatch between core health and peripheral pathology is reflected in the matrix as specific divergence signatures. The system identifies this inconsistency between core and peripheral response patterns as important evidence of concealed behavior.

[0058] The system statistically analyzes the cumulative number and severity of logical deviations across various psychological dimensions for respondents, constructing a logical conflict feature vector. Each component of the feature vector corresponds to a psychological dimension, and the component value reflects the conflict intensity of that dimension. The feature vector has approximately 10 to 15 dimensions, depending on the number of psychological dimensions involved in the scale system. Common scale systems include dimensions such as depression, anxiety, obsessive-compulsive disorder, interpersonal relationships, hostility, fear, paranoia, psychoticism, and somatization. Furthermore, the system calculates the global conflict density, defined as the ratio of the number of logical deviations to the total number of tested item pairs, reflecting the overall logical consistency level of the respondents. The global conflict density serves as an additional dimension of the feature vector. The logical conflict feature vector, as a quantitative representation of the respondents' semantic-layer masking features, is output to the subsequent fusion step. This step and step S2 run in parallel and independently after the multimodal assessment dataset is constructed in step S1. They extract masking features from the semantic logic layer and behavioral dynamics layer respectively, without interdependence. The two feature paths are fused in step S4.

[0059] S4 receives the behavioral fluctuation entropy feature vector and the logical conflict feature vector, and uses a dual-path parallel feature extraction network and a self-attention fusion mechanism to perform a comprehensive evaluation, and outputs the corrected crisis assessment result. This step aims to non-linearly fuse the behavioral features extracted in S2 with the logical features extracted in S3 to overcome the limitations of single-dimensional judgment. Different respondents employ different concealment strategies; some meticulously disguise their answers but reveal flaws in their behavior, while others demonstrate good behavioral control but struggle to maintain logical consistency, and still others exhibit anomalies in both dimensions. Simple fixed-weight fusion cannot accommodate these individual differences; a deep attention mechanism is needed to extract the most essential warning features from multi-dimensional heterogeneous data.

[0060] The system receives a behavioral fluctuation entropy feature vector from S2 and a logical conflict feature vector from S3. The two feature vectors characterize the respondents' masking features from the behavioral dynamics layer and the semantic logic layer, respectively, but they differ in dimensionality and scale, requiring nonlinear mapping and collaborative analysis through a deep attention fusion network.

[0061] The system constructs a dual-path parallel feature extraction network. The first path is the behavioral dynamics branch, and the second path is the semantic conflict branch. The two branches process their respective input features in parallel, and then synthesize them through an attention fusion mechanism.

[0062] The Behavioral Dynamics branch receives the convolutional feature sequence output from S2 as input. This sequence preserves the complete temporal structure of the question-answering process, with each time step corresponding to the behavioral feature representation of one question, and the dimension being the number of questions multiplied by 128. This branch utilizes a recurrent neural network to extract the long-range dependency features of behavioral fluctuation entropy evolving over time. Temporal convolutional networks excel at capturing local temporal patterns, while recurrent neural networks excel at modeling global long-range dependencies. The combination of the two can comprehensively characterize the dynamic evolution of masking behavior at different time scales. The system uses a Long Short-Term Memory (LSTM) network as the basic unit of the recurrent neural network. LSTM, through three gating mechanisms—input gate, forget gate, and output gate—can effectively capture long-range dependencies in temporal data and avoid the gradient vanishing problem. The system inputs the convolutional feature sequence into the LSTM network in the order of question-answering time. The hidden layer dimension of the LSTM network is set to 64, and the feature extraction capability is enhanced through a two-layer stacked structure. After LSTM processing, the system extracts the hidden state of the last time step as the long-range temporal feature representation of behavioral dynamics, with a dimension of 64. Then, feature transformation is performed through two fully connected neural networks. The first fully connected layer maps temporal features to a higher-dimensional hidden space. The default number of hidden layer nodes is set to 64, and the ReLU activation function is used to introduce non-linearity. The second fully connected layer compresses the features into a unified representation space, and the output dimension is set to 32 by default. The output of the behavioral dynamics branch is the high-level feature vector of the behavioral layer, with a dimension of 32.

[0063] The semantic conflict branch receives the logical conflict feature vector output from S3 as input. This vector contains the conflict density of each psychological dimension and the global conflict density, with dimensions ranging from approximately 10 to 15. This branch utilizes a graph attention network to extract the spatial features of the conflict detection matrix, capturing logical violations between different dimensions. The system uses the conflict intensity of each dimension in the logical conflict feature vector as node features to construct a psychological dimension association graph. This graph differs from the question-level knowledge graph in S1, where the nodes are questions, focusing on micro-semantic relationships between questions; while the nodes in this dimension association graph are psychological dimensions, such as depression, anxiety, and interpersonal relationships, focusing on macro-clinical relationships between dimensions. By modeling the graph structure at different levels of abstraction, the system can extract dimension-level conflict patterns from question-level logical conflicts, capturing cross-dimensional masking strategy features. The edges between nodes in the graph represent clinical relationships between dimensions; for example, the depression dimension and the sleep quality dimension are usually strongly correlated, and the anxiety dimension and the somatization dimension often co-occur. The system uses a single-layer graph attention network to aggregate features from this dimension association graph. The graph attention layer calculates the attention weights between each dimension node and its related dimension nodes, aggregating the conflict intensity information of neighboring nodes through attention weighting to obtain an enhanced feature representation containing inter-dimensional correlations. The output dimension of the graph attention layer is set to 32. Then, feature transformation is performed through two fully connected neural network layers. The first fully connected layer maps graph features to the hidden space, with a default hidden layer node count of 64 and a ReLU activation function. The second fully connected layer compresses the features into a unified representation space, with an output dimension of 32 (default value), consistent with the dimension of the high-level feature vectors of the behavior layer, ensuring symmetry in subsequent fusion. The output of the semantic conflict branch is the high-level feature vector of the logic layer, with a dimension of 32.

[0064] The system introduces a self-attention mechanism for adaptive fusion of dual-path features. This attention mechanism adaptively determines the contribution of each dimension feature to the final judgment based on the actual distribution of the input features, achieving a personalized fusion strategy. If the system detects extremely high behavioral entropy in the respondent, it indicates that their current answering state is extremely unstable; in this case, the model will automatically increase the weight of behavioral features. If the behavioral performance is stable but logical conflicts are frequent, it indicates that the respondent is a skilled concealer; in this case, the model will increase the weight of logical conflict features.

[0065] The attention fusion module comprises two gating units: a behavioral feature gate and a logical feature gate. The gating mechanism borrows from the design concept of gated recurrent units, achieving dynamic feature selection through learnable gating weights. The calculation process for the behavioral feature gate is as follows: The high-level feature vectors from the behavioral layer and the logical layer are concatenated to obtain a joint feature vector of dimension 64; this joint feature vector is then input into a fully connected layer with an output dimension of 32, and mapped to the interval between 0 and 1 using a sigmoid activation function to obtain the behavioral feature gating weight vector; each element of the behavioral feature gating weight vector controls the proportion of the corresponding dimension of the high-level feature vectors from the behavioral layer passing through. Similarly, the calculation process for the logical feature gate is as follows: The concatenated joint feature vector is input into another fully connected layer with an output dimension of 32, and activated using a sigmoid function to obtain the logical feature gating weight vector. The calculation of the gating weights is essentially an attention allocation mechanism, automatically determining the importance of each feature by learning the interaction relationships between input features.

[0066] The system multiplies the high-level feature vector of the behavioral layer element-wise with the behavioral feature gating weight vector to obtain the gated modulated behavioral features. It then multiplies the high-level feature vector of the logic layer element-wise with the logic feature gating weight vector to obtain the gated modulated logic features. Finally, the two gated modulated feature vectors are added element-wise to obtain a fused feature vector with a dimension of 32. This fused feature vector contains information from both the behavioral and logic layers, and the fusion weights are adaptively learned by the gating unit based on the input features; different inputs will produce different fusion weight distributions.

[0067] The decision output layer receives the fused feature vector and maps it to a masking score through two fully connected layers. The first fully connected layer maps the fused features to an intermediate representation, with a default of 16 nodes and a ReLU activation function. The second fully connected layer maps the intermediate representation to a single scalar output, which is then constrained to the range of 0 to 1 by a sigmoid activation function, resulting in the masking score. A masking score close to one indicates that the respondent is engaging in highly deceptive behavior, with their scale score significantly deviating from their true psychological state and having extremely low credibility. A masking score close to zero indicates that the respondent is responding honestly, and the scale score has high credibility.

[0068] The training of deep attention fusion networks employs a multi-stage strategy, making full use of both limited labeled data and a large amount of unlabeled data. Training is divided into three stages: a contrastive learning pre-training stage, a supervised fine-tuning stage, and a posterior validation enhancement stage.

[0069] The contrastive learning pre-training phase utilizes a large amount of unlabeled historical assessment data for self-supervised learning. The system treats multiple assessment records completed by the same respondent at different times as positive sample pairs. If a respondent completes the same or similar scales in two assessments three months apart, the system extracts the behavioral fluctuation entropy feature vector and logical conflict feature vector from these two assessments, respectively, and encodes them into two representation vectors through a fusion network. The goal of contrastive learning is to maximize the similarity of positive sample pairs while minimizing the similarity between different respondents. The system employs a contrastive loss function, requiring the cosine similarity of the representation vectors of positive sample pairs to be as close to one as possible. For negative sample pairs, the cosine similarity of the representation vectors is required to be as close to zero as possible. Through contrastive learning, the fusion network learns a representation space sensitive to the consistency of individual response patterns, enabling it to identify the inherent correlation between the behavioral and logical features of the same individual at different times. The default size of the historical assessment data used in the contrastive learning phase is 5000 to 10000 records, the number of training iterations is 50 to 100 epochs, and the default learning rate is 0.001.

[0070] The supervised fine-tuning phase utilizes a small number of expert-annotated masking and honest samples to fine-tune the pre-trained fusion network. Expert-annotated data sources include: in subsequent psychological interviews, some respondents admitted to masking their actions during the assessment; these cases were recorded by the counselor, and their assessment data were labeled as masking samples; for respondents who showed high cooperation, consistency in their statements, and no obvious defensiveness during the interviews, their assessment data were labeled as honest samples; 3 to 5 senior psychology experts were invited to independently blind-review and annotate 100 to 200 typical cases. Experts comprehensively judged whether masking behavior existed based on the answer patterns, behavioral characteristics, and logical consistency in the assessment data. A majority voting mechanism was used to resolve disagreements among experts, and only samples with expert consensus reaching a preset consensus threshold were included in the training set to ensure the reliability and objectivity of the labeled data. The default number of labeled samples was 200 to 300, with masking samples accounting for 30 to 40, because honest respondents are the majority in real-world scenarios. The fine-tuning phase uses a binary classification cross-entropy loss function, with the optimization objective being to minimize the difference between the predicted masking score and the true label. The Adam optimizer is used, with the default learning rate reduced to 0.0001 to avoid overfitting. Fine-tuning iterations are performed for 20 to 50 epochs. To address the issue of insufficient labeled samples, the system employs data augmentation strategies, such as adding Gaussian noise to feature vectors and performing random dropout, to increase the diversity of training samples.

[0071] The post-validation enhancement phase continuously optimizes the model using post-validation feedback data accumulated in real-world applications. After system deployment, for cases flagged as high-risk and confirmed through manual intervention, as well as missed cases deemed low-risk but subsequently experiencing psychological crises, the system uses these cases as weakly supervised signals for incremental learning. Specifically, for cases confirmed to involve concealment, the system sets the true label of the concealment score to 1; for cases confirmed to involve honest responses, the label is set to 0; and for missed cases, the label weights are adjusted based on the severity of the subsequent crisis. The system employs online learning or small-batch incremental updates, periodically fine-tuning the parameters of the fusion network using accumulated post-validation data to achieve continuous model evolution. The post-validation enhancement uses a small learning rate, with a default value of 0.00005, and uses 10 to 50 new samples per update to avoid catastrophic forgetting of historical learning outcomes.

[0072] The system corrects the original scale assessment results based on the concealment score, outputting a crisis authenticity correction model. The original scale assessment results are the respondents' scores on each psychological dimension and their overall risk level. When the concealment score exceeds a preset concealment judgment threshold, the system determines that the respondent has engaged in concealment behavior, requiring correction of the assessment results. The concealment judgment threshold is determined based on performance analysis on the validation dataset, with a default value set to 0.7, which strikes a balance between detection sensitivity and false positive rate.

[0073] The system pre-labels the scoring direction attributes for each psychological dimension in the scale configuration file, dividing them into negative and positive scoring dimensions. Negative scoring dimensions refer to dimensions where low scores represent health and high scores represent severe problems, such as crisis-related dimensions like depression level, anxiety level, and intensity of negative thoughts. Masking behavior in these dimensions manifests as respondents avoiding high-risk options, scores being artificially suppressed, and the true crisis state being underestimated. Positive scoring dimensions refer to dimensions where high scores represent health or a positive state, such as positive resource dimensions like self-efficacy, social support, and sense of meaning in life. Masking behavior in these dimensions manifests as respondents exaggerating their positive state, scores being artificially inflated, and the true functioning level being overestimated.

[0074] The fusion network ultimately outputs a truthfulness coefficient between 0 and 1, which is the complement of the concealment score. This coefficient is used to correct the total score of the original scale. The specific implementation of the correction logic is as follows: the system adopts different correction strategies based on the scoring direction attribute of each dimension. For negatively scored dimensions, the system calculates a correction coefficient, defined as 1 plus the product of the concealment score and the scaling factor. If the correction coefficient is greater than 1, the original score is adjusted upwards to reflect the underestimated true crisis level of the respondents. For example, if the original score is low, but the truthfulness coefficient is extremely low, indicating high concealment, the system will compensate the score in reverse, raising the warning level. For positively scored dimensions, the correction coefficient is defined as 1 minus the product of the concealment score and the scaling factor. If the correction coefficient is less than 1 and not lower than 0, the original score is adjusted downwards to eliminate the respondents' falsely reported positive state. The scaling factor controls the strength of the correction and is set according to the risk tolerance of the application scenario; the default value is 0.8. For high-risk scenarios such as crisis risk assessment, the scaling factor can be set to 1.0 to improve sensitivity; for general psychological state assessment, the scaling factor can be set to 0.5 to avoid over-correction. The corrected dimension scores are obtained by multiplying the scores of each dimension of the original scale by the corresponding correction coefficient.

[0075] The system recalculates the overall crisis score based on the corrected dimensional scores. The overall crisis score is derived from the dimensional scores through weighted summation or a decision tree model, depending on the scale system and assessment criteria used. The system outputs the corrected crisis assessment results, including the corrected dimensional scores, the overall crisis score, and the concealment score. These results are then passed to the early warning report generation step.

[0076] S5, based on the revised crisis assessment results, uses risk level determination rules and dynamic evolution prediction strategies to generate an in-depth early warning report containing risk profiles and intervention recommendations.

[0077] This step transforms abstract feature vectors into decision support information usable by mental health counselors, addressing the pain points of inaccurate early warnings and inability to intervene using existing technologies. The final output is not just a score, but a risk profile containing behavioral logic evidence, providing clear technical guidance for subsequent human intervention.

[0078] The system receives the revised crisis assessment results output by S4, including the revised comprehensive crisis score and concealment score. Based on the comprehensive crisis score, it classifies the risk level and predicts its dynamic evolution. The system has a pre-set four-level risk level system: Red (extremely high risk), Orange (high risk), Yellow (medium risk), and Blue (concern). The rules for determining the risk level are stored in a risk level rule base, which is based on statistical analysis of a large amount of historical warning data and risk assessment standards from clinical psychology.

[0079] The static grading rule for risk level determination is as follows: the comprehensive crisis score is divided into four intervals, each corresponding to a risk level. The extremely high-risk level corresponds to a score exceeding the first threshold, indicating that the respondent is experiencing a severe psychological crisis, may have negative thoughts or self-harming behavior, and requires immediate crisis intervention. The high-risk level corresponds to a score between the second and first thresholds, indicating that the respondent has significant psychological distress, such as moderate to severe depression or anxiety, and requires prompt psychological counseling. The moderate-risk level corresponds to a score between the third and second thresholds, indicating that the respondent has some psychological stress or adaptation problems, and psychological counseling is recommended. The attention level corresponds to a score below the third threshold, indicating that the respondent's overall mental state is healthy and can be included in the scope of routine attention. The specific values ​​of each threshold are determined through analysis of historical data and receiver operating characteristic curves, and may vary in different psychological dimensions and application scenarios.

[0080] After classifying respondents into corresponding risk levels, the system performs dynamic trend prediction. The system checks for historical assessment records for each respondent. For respondents with historical records, the system extracts feature vectors from their past assessments, including historical logical conflict features and behavioral fluctuation entropy features. Combining the respondent's past assessment records, the system calculates the spatial migration trajectory of the crisis feature vector. The system compares the feature vector of the current assessment with the feature vector of the most recent historical assessment, calculating the Euclidean distance in the feature space as a measure of the magnitude of psychological state change. The system calculates the feature vectors of the three most recent assessments, fits a time evolution trajectory, and calculates the slope of the trajectory as a measure of the speed of psychological state change.

[0081] If the rate of change in psychological state exceeds a preset deterioration threshold, the system triggers a dynamic early warning mechanism. If the rate of change accelerates, even if the current level is at the yellow medium-risk level, early intervention recommendations will be triggered. Even if the respondent's current overall crisis score is at a low-risk level, if their psychological state is rapidly deteriorating, the system will raise the warning level or add a special note to the report, advising mental health counselors to pay attention in advance. The dynamic early warning mechanism can capture early signs of psychological crisis, providing warnings before the crisis fully manifests, thus improving the foresight of early warnings.

[0082] The system generates risk profiles and intervention indexes for respondents. The risk profile is a multi-dimensional descriptive analysis of the respondent's psychological state, comprising three core components.

[0083] The first part explains the criteria for determining concealment. The system explicitly identifies whether the respondent is engaging in concealment behavior. If the concealment score exceeds the threshold for determining concealment, the system marks the respondent as exhibiting significant concealment behavior in the report and further analyzes the main dimensions of concealment. By retrospectively analyzing the specific conflict events recorded in the S3 logical conflict detection step, the system clearly identifies which dimensions the respondent exhibits significant behavioral inconsistencies or logical contradictions. For example, if the respondent has numerous conflicts in the item pairs related to the depression dimension, but fewer conflicts in the anxiety and interpersonal relationship dimensions, it is determined that they are primarily concealing depressive symptoms. The system also analyzes the types of concealment strategies, distinguishing between core concealment and peripheral concealment. Core concealment refers to the respondent appearing healthy in core crisis symptoms such as negative thoughts, but revealing problems in peripheral symptoms such as somatization or interpersonal relationships. This mismatch pattern suggests that they may be concealing in the core dimensions. Peripheral concealment refers to the respondent acknowledging problems in core symptoms but overly romanticizing their general mental state.

[0084] The second part focuses on core symptom localization. Based on the S3 conflict matrix analysis, the system strips away the respondents' deliberately embellished aspects, focusing on their true psychological distress. For respondents exhibiting concealment behavior, the system identifies which dimensions, while scoring low, show obvious logical contradictions or behavioral abnormalities, inferring that these dimensions may be the source of their true distress. Combining the revised dimension scores with concealment pattern analysis, the system identifies core symptoms, such as masked depression, high anxiety with social avoidance, and adjustment disorder. Core symptom localization provides mental health counselors with clear intervention directions.

[0085] The third part outputs targeted intervention strategies. Based on the identified specific risk factors, the system indexes the most suitable interview strategy from a pre-set intervention database. This intervention strategy knowledge base, built by experts in the field of psychological counseling, contains intervention methods and communication skills for different psychological problems and individual characteristics. For highly concealing respondents, such as highly concealing students, the knowledge base emphasizes using indirect communication techniques, suggesting counselors adopt indirect communication strategies to avoid directly asking sensitive questions in the initial interview, instead gradually guiding the respondent to open up by building trust. For rapidly deteriorating respondents, the knowledge base emphasizes shortening the intervention period, increasing the frequency of interviews, and closely monitoring changes in their psychological state. The system presents the retrieved intervention strategies in structured text form in the warning report to assist psychological counselors in developing personalized intervention plans.

[0086] The system integrates risk level, revised crisis score, concealment score, risk profile description, and intervention strategy recommendations into a structured early warning report. The report format includes: a basic information section for the respondent, listing their name, student ID, college, and assessment date; a risk assessment section, using prominent colors to indicate risk levels and listing the overall crisis score and scores for each dimension; a concealment analysis section, explaining the concealment score, indicating the presence of concealment behavior, and analyzing the main dimensions and strategy types of concealment; a risk profile section, describing core symptoms and psychological state characteristics; and an intervention recommendation section, listing matching intervention strategies and precautions. For respondents with historical records, the report also includes a dynamic trend analysis section, showing the evolution trajectory and rate of change of their psychological state.

[0087] The system pushes the generated warning reports to the mental health counselor's work platform. For respondents at extremely high risk, the system also triggers an instant notification mechanism, sending an emergency warning to the on-duty staff at the mental health counseling center via SMS or app push to ensure timely response. For respondents at other risk levels, the reports are summarized in a pending list on the work platform, allowing mental health counselors to view and process them in batches according to risk level.

[0088] In a psychological screening of freshmen at a university in the fall semester of the 2025 academic year, this technical solution was used to provide psychological crisis early warning for 3,000 freshmen. Taking student A as an example, this student completed a total of 90 questions on the Depression Self-Rating Scale, Anxiety Self-Rating Scale, and Symptom Self-Rating Scale during the assessment process. The following is an illustrative case demonstration based on a real-world deployment scenario. The specific values ​​involved are exemplary data, and the values ​​in actual applications may vary due to individual differences and the assessment environment. This is only used to illustrate the implementation process of the technical solution and does not constitute a limitation on the scope of protection.

[0089] The system collected student A's answer data and interaction behavior data in real time. Table 1 shows examples of student A's answers to some questions, and Table 2 shows examples of student A's interaction behavior characteristics to some questions.

[0090] Table 1. Example of Student A's Answer Note: Option scores range from 1 to 5, where 1 indicates never or very minor, and 5 indicates always or very serious. Reverse-scoring questions have been converted by the system.

[0091] Table 2, Examples of Student A's Interactive Behavior Characteristics Table 1 shows that Student A consistently scored the lowest marks on crisis-sensitive questions directly related to low mood and negative thoughts, seemingly indicating a very healthy psychological state. However, they scored a relatively high mark (4 points) on life satisfaction questions, suggesting significant dissatisfaction with life. Table 2 shows that Student A's answering time, number of option switching attempts, and hesitation trajectory length on crisis-sensitive questions Q015 and Q023 were all higher than those on conventional assessment questions Q056 and Q078, suggesting that they underwent complex cognitive processing on these questions.

[0092] The system first constructed a high-dimensional fingerprint feature database of the respondent's interactive behavior. The behavioral data of the routine assessment questions Q056 and Q078, with an answering time of about 3 seconds and very few switching, constituted the individualized behavioral fingerprint baseline of the respondent.

[0093] The system performs cognitive load analysis on student A's interactive behavior data, extracting behavioral fluctuation entropy features. It calculates the average answer time and standard deviation for routine assessment questions, using this as a baseline for individual behavioral fingerprints. The system also calculates the average answer time for crisis-sensitive questions; the standardized deviation score from the baseline exceeds the normal range, indicating anomalies. The system inputs the answer behavior sequence into a temporal convolutional network for deep feature extraction, identifying local abnormal patterns in student A's answer rhythm for crisis-sensitive questions. The answer time for crisis-sensitive questions is discretized into intervals, and Shannon entropy is calculated; the result is higher than the baseline entropy level of the normal group. The overall increase in behavioral entropy suggests that student A has abnormal cognitive load on these questions, contrasting sharply with the abnormal behavior on crisis-sensitive questions. This abnormal cognitive load in answering behavior suggests that student A consumed significant cognitive resources during the concealment process, consistent with the expectations of the two-way processing theory.

[0094] The system performed cross-scale logical conflict detection on Student A's answers. The system invoked a pre-built knowledge graph in S1. A logical support relationship existed between question Q023 (negative thoughts) and question Q056 (life satisfaction), indicating that individuals dissatisfied with their lives are more likely to have negative thoughts. Student A selected "never have negative thoughts" for Q023 (score 1 point) and showed significant dissatisfaction with life for Q056 (score 4 points). The system converted the answers to both questions into semantic vectors, processed them using a graph attention network, and calculated the attention-weighted similarity. The result was below the expected threshold for logical support, and the system determined it as a logical deviation. Similarly, the system detected logical deviations in multiple pairs of items for Student A. The constructed logical conflict detection matrix and logical conflict feature vectors showed that the conflict density in the depression dimension was significantly higher than the baseline level of the normal group.

[0095] The system inputs the behavioral fluctuation entropy feature vector output by S2 and the logical conflict feature vector output by S3 into a deep attention fusion network. S2, in addition to outputting a fixed-dimensional feature vector, also retains the convolutional feature sequence of the temporal convolutional network. The LSTM network of the behavioral dynamics branch processes this convolutional feature sequence, capturing the long-term evolution pattern of Student A's answering behavior at different time scales and extracting temporally dependent features. The logical conflict features are extracted at the psychological dimension level through the graph attention network of the semantic conflict branch, revealing cross-dimensional conflict pattern features. The system calculates the gating fusion weights of the two feature paths through a self-attention mechanism. Since Student A exhibits abnormalities in both the behavioral and logical layers, the fusion network adaptively assigns higher weights to both feature paths. When the masking score output by the fusion network through the gating mechanism exceeds a preset masking threshold, the system determines that Student A is engaging in masking behavior.

[0096] The system corrected Student A's original scale assessment results. The original depression dimension score was below the clinical threshold for mild depression, which, according to traditional assessment, falls within the healthy range. The system calculated a correction coefficient based on the concealment score. Since the depression dimension is a negatively scored dimension, the correction coefficient was greater than 1, thus increasing the original score. The corrected depression dimension score exceeded the mild depression threshold, and the risk level was upgraded from blue (concern) to yellow (moderate risk).

[0097] The system generated a warning report for Student A, with a risk level of yellow (medium risk) and a concealment score exceeding the concealment threshold. The risk profile indicated that the respondent exhibited significant concealment during the assessment, primarily displaying a false sense of health in core crisis dimensions such as depressive mood and negative thoughts (selecting the lowest score in both Q015 and Q023), but admitting significant problems in peripheral dimensions such as life satisfaction (score of 4 in Q056), presenting a core concealment pattern. Based on the revised score, it was inferred that the respondent may have a hidden depressive tendency. It was recommended that the counselor use a non-direct communication strategy, gradually exploring the respondent's true psychological state after establishing a trusting relationship, avoiding directly asking sensitive questions such as negative thoughts during the initial interview.

[0098] Based on the early warning report, the psychological counseling center focused its attention on Student A. In subsequent one-on-one interviews, the counselor employed the communication strategies recommended by the system. Through three interviews, a trusting relationship was gradually established, and Student A finally confessed that he was indeed worried that truthful answers during the assessment would affect his scholarship evaluation, and therefore deliberately concealed his true emotional distress. The counselor provided timely psychological support and cognitive-behavioral intervention, effectively helping Student A improve his mental state. This case verifies that the technical solution of this invention can effectively identify concealment behavior and improve the accuracy of psychological crisis early warning.

[0099] In this survey, the system identified several students exhibiting concealment behavior. After adjustments, the risk level of some students was upgraded to high-risk or very high-risk, and all received timely psychological intervention. Compared with the assessment results using traditional early warning systems, the technical solution of this invention performed well in terms of the accuracy of concealment behavior identification and the ability to detect high-risk individuals. It effectively reduced the risk of missed detection by traditional methods, improved the practicality and reliability of the early warning system, and provided strong technical support for psychological crisis prevention and control in universities.

[0100] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-scale psychological crisis early warning method, characterized in that, Includes the following steps: S1. Obtain psychological assessment task requests, construct a cross-scale knowledge graph and train it through a graph attention network, while simultaneously collecting high-dimensional fingerprints of respondents' answers and interactive behaviors in real time to obtain a cross-scale knowledge graph and a multimodal assessment dataset containing the original assessment results; the cross-scale knowledge graph is constructed by identifying question pairs and labeling logical association types through semantic similarity calculation, with questions as nodes and logical association relationships as edges. S2, acquire a multimodal assessment dataset, and extract cognitive load shift features using a combination of temporal convolutional networks and information entropy calculation, outputting a behavioral fluctuation entropy feature vector and a convolutional feature sequence; the behavioral fluctuation entropy feature vector is constructed by concatenating the behavioral depth feature vector extracted by the temporal convolutional network with the behavioral entropy value statistical feature vector calculated based on the Shannon entropy formula; S3. Based on the knowledge graph and the answer content, a graph attention network is used to perform cross-scale semantic consistency analysis, generate a logical conflict detection matrix and output a logical conflict feature vector. The logical conflict feature vector is constructed by determining whether there is a logical deviation in the question pair based on the logical association type and the preset consistency threshold, and by counting the number and severity of deviations in each psychological dimension. S4 receives the behavioral fluctuation entropy feature vector and the logical conflict feature vector, and uses a dual-path parallel feature extraction network and a self-attention fusion mechanism to perform a comprehensive evaluation, and outputs the corrected crisis assessment result. S5, based on the revised crisis assessment results, uses risk level determination rules and dynamic evolution prediction strategies to generate an in-depth early warning report containing risk profiles and intervention recommendations.

2. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S1 includes: A pre-trained language model was used to semantically vectorize the text of all psychological scale items. Semantic similarity calculation was used to identify semantically related item pairs. Experts in the field of clinical psychology reviewed and labeled the item pairs to determine the logical relationship type between them. The logical relationship type includes logical support relationship and logical conflict relationship, and the association strength weight was labeled. The labeled logical relationships are constructed into a cross-scale knowledge graph, where nodes represent questions and edges represent logical relationships between questions. The training of the graph attention network includes: constructing a graph attention network on the cross-scale knowledge graph, using a multi-head attention mechanism to calculate the attention weights between nodes and their neighbors, and using a contrastive learning framework to train the graph attention network with historical test data, so that the graph attention network learns question representations that are sensitive to the consistency of individual response patterns.

3. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S1 further includes: When respondents take the test, physical layer features, dynamic layer features and contextual features are collected to form a high-dimensional fingerprint of the interaction behavior. The physical layer features include answer time, visual processing time, hesitation trajectory distance and backtracking correction rate. The dynamic layer features include the coordinate sequence of option switching trajectory and mouse movement trajectory, as well as speed and acceleration changes. The collected raw behavioral data were normalized to generate a behavioral fingerprint baseline for each respondent. The answer content data and interaction behavior data were correlated and integrated to construct a multimodal assessment dataset. Each record in the dataset contains respondent identifier, question identifier, option score, question type and multidimensional high-dimensional fingerprint features. The question types include crisis sensitivity type and routine assessment type.

4. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S2 includes: The response time, first response delay, number of option switching times, and hesitation trajectory length of each question were extracted from the multimodal assessment dataset. Based on the question type, the questions were divided into regular assessment questions and crisis-sensitive questions. We extracted the behavioral characteristics of respondents on routine assessment questions, and calculated the mean and standard deviation of answer time, first response delay, number of option switching and hesitation trajectory length to establish an individualized behavioral fingerprint baseline. For each crisis-sensitive question, we calculated the standardized deviation score of its behavioral characteristics relative to the individualized behavioral fingerprint baseline, which was defined as the cognitive load increment.

5. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S2 further includes: The response time sequence, response delay sequence, switching number sequence, and trajectory length sequence of respondents are used as multi-channel inputs. A one-dimensional temporal convolutional network with multi-scale convolutional kernels is constructed to extract deep features. The behavior depth feature vector is obtained by global max pooling, while the convolutional feature sequence is preserved. The response time of respondents on crisis-sensitive questions is discretized into intervals, a discrete probability density function is constructed, and the duration entropy is calculated based on the Shannon entropy formula. The entropy values ​​of the number of option switching and the first response delay are calculated separately to obtain the comprehensive behavioral entropy value. The behavioral depth feature vector is concatenated with the statistical feature vector containing the comprehensive behavioral entropy value, duration entropy, switching entropy, delay entropy and deviation mean to construct the behavioral fluctuation entropy feature vector.

6. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S3 includes: For each completed question, the question text and the text of the options selected by the respondent are concatenated to form a combined text, which is then input into a pre-trained language model to obtain the response content vector. The response content vector is used as the dynamic initial feature of the corresponding node in the cross-scale knowledge graph, and the feature is propagated and updated through a graph attention network to obtain the response context representation vector for each question. Iterate through all logically related question pairs, calculate the attention-weighted similarity between the response context representation vectors, and determine logical deviations when the similarity is lower than the preset positive consistency threshold for logically supported question pairs, and logical deviations when the similarity is higher than the preset negative consistency threshold for logically conflicting question pairs. Calculate the cumulative value of the number of logical deviations and the severity of deviations in each psychological dimension, calculate the global conflict density, and construct a logical conflict feature vector.

7. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S4 includes: The dual-path parallel feature extraction network includes a behavioral dynamics branch and a semantic conflict branch. The behavioral dynamics branch receives the convolutional feature sequence, uses a long short-term memory network to extract long-range temporal feature representations, and then performs feature transformation through a fully connected neural network to output a high-level feature vector of the behavioral layer. The semantic conflict branch receives the logical conflict feature vector, constructs a psychological dimension association graph by using the conflict intensity of each psychological dimension as node features, performs feature aggregation using a graph attention network, and then performs feature transformation through a fully connected neural network to output the high-level feature vector of the logical layer. The self-attention fusion mechanism includes two gating units: a behavioral feature gate and a logical feature gate. The high-level feature vectors of the behavioral layer and the high-level feature vectors of the logical layer are concatenated and then input into the corresponding fully connected layers to obtain the gating weight vectors. Each high-level feature vector is multiplied element-wise with the corresponding gating weight vector and then added element-wise to obtain the fused feature vector.

8. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S4 further includes: The fused feature vector is mapped to a concealment score through a fully connected network. The concealment score ranges from 0 to 1. When the concealment score exceeds a preset concealment judgment threshold, it is determined that the respondent has concealed behavior. Different correction strategies are adopted according to the scoring direction attribute of the psychological dimension. For negative scoring dimensions, the correction coefficient is set to be greater than 1 to adjust the original score, and for positive scoring dimensions, the correction coefficient is set to be less than 1 to adjust the original score. The original scale's dimension scores are multiplied by their corresponding correction coefficients to obtain the corrected dimension scores. The overall crisis score is then recalculated based on these corrected dimension scores.

9. The multi-scale psychological crisis early warning method according to claim 1, characterized in that, S5 includes: A four-level risk rating system is pre-set, which divides the range of the comprehensive crisis score into four intervals. The score is determined to be extremely high risk if it exceeds the first pre-set threshold, high risk if it is between the second pre-set threshold and the first pre-set threshold, medium risk if it is between the third pre-set threshold and the second pre-set threshold, and at least at concern level if it is below the third pre-set threshold. The dynamic evolution prediction strategy includes: extracting feature vectors from respondents' historical assessments, calculating the Euclidean distance between the current assessment feature vector and the historical assessment feature vector in the feature space as the magnitude of psychological state change, fitting the time evolution trajectory to calculate the trajectory slope as the speed of psychological state change, and triggering a dynamic early warning mechanism when the speed of change exceeds a preset deterioration speed threshold; the risk profile includes explanations of concealment, core symptom localization, and output of targeted intervention strategies.

10. A multi-scale psychological crisis early warning system, used to perform the steps in the multi-scale psychological crisis early warning method as described in any one of claims 1-9, characterized in that, include: The graph construction and acquisition module is used to obtain psychological assessment task requests, construct a cross-scale knowledge graph and train it through a graph attention network, and simultaneously collect high-dimensional fingerprints of respondents' answers and interaction behaviors in real time to obtain a cross-scale knowledge graph and a multimodal assessment dataset containing the original assessment results. The cross-scale knowledge graph is constructed by identifying question pairs and labeling logical association types through semantic similarity calculation, with questions as nodes and logical association relationships as edges. The cognitive load shift feature extraction module is used to acquire a multimodal assessment dataset. It uses a combination of temporal convolutional networks and information entropy calculation to extract cognitive load shift features and outputs a behavioral fluctuation entropy feature vector and a convolutional feature sequence. The behavioral fluctuation entropy feature vector is constructed by concatenating the behavioral depth feature vector extracted by the temporal convolutional network with the behavioral entropy value statistical feature vector calculated based on the Shannon entropy formula. The cross-scale semantic consistency analysis module is used to perform cross-scale semantic consistency analysis based on the knowledge graph and the answer content, using a graph attention network to generate a logical conflict detection matrix and output a logical conflict feature vector. The logical conflict feature vector is constructed by determining whether there is a logical deviation in the question pair based on the logical association type and the preset consistency threshold, and by counting the number and severity of deviations in each psychological dimension. The dual-path feature fusion and comprehensive evaluation module receives behavioral fluctuation entropy feature vectors and logical conflict feature vectors, and uses a dual-path parallel feature extraction network and self-attention fusion mechanism to perform comprehensive evaluation, outputting a corrected crisis assessment result. The in-depth early warning report generation module is used to generate in-depth early warning reports containing risk profiles and intervention recommendations based on the revised crisis assessment results, risk level determination rules, and dynamic evolution prediction strategies.

Images